The following is a conversation with Ishan Mizra,
research scientist at Facebook AI Research,
who works on self-supervised machine learning
in the domain of computer vision,
or in other words, making AI systems understand
the visual world with minimal help from us humans.
Transformers and self-attention has been successfully used
by OpenAI GPT-3 and other language models
to do self-supervised learning in the domain of language.
Ishan, together with Yan Likun and others,
is trying to achieve the same success
in the domain of images and video.
The goal is to leave a robot watching YouTube videos
all night, and in the morning,
come back to a much smarter robot.
I read the blog post self-supervised learning,
The Dark Matter of Intelligence by Ishan and Yan Likun,
and then listened to Ishan's appearance
on the Excellent Machine Learning Street Talk podcast.
And I knew I had to talk to him.
By the way, if you're interested in machine learning and AI,
I cannot recommend the ML Street Talk podcast highly enough.
Those guys are great.
Quick mention of our sponsors.
On it, the information, Grammarly, and Athletic Greens.
Check them out in the description to support this podcast.
As a side note, let me say that for those of you
who may have been listening for quite a while,
this podcast used to be called
Artificial Intelligence podcast,
because my life passion has always been,
will always be artificial intelligence,
both narrowly and broadly defined.
My goal with this podcast is still
to have many conversations with world-class researchers
in AI, math, physics, biology, and all the other sciences.
But I also want to talk to historians, musicians, athletes,
and of course, occasionally comedians.
In fact, I'm trying out doing this podcast
three times a week now
to give me more freedom with guest selection
and maybe get a chance to have a bit more fun.
Speaking of fun, in this conversation,
I challenged the listener to count the number of times
the word banana is mentioned.
Ishan and I used the word banana as the canonical example
at the core of the hard problem of computer vision
and maybe the hard problem of consciousness.
This is the Lex Friedman podcast
and here is my conversation with Ishan Mizra.
What is self-supervised learning?
And maybe even give the bigger basics
of what is supervised and semi-supervised learning.
And maybe why is self-supervised learning
a better term than unsupervised learning?
Let's start with supervised learning.
So typically for machine learning systems,
the way they're trained is you get a bunch of humans.
The humans point out particular concepts.
So if it's in the case of images,
you want the humans to come and tell you
what is present in the image,
draw boxes around them,
draw masks of things, pixels,
which are of particular categories or not.
For NLP, again, there are lots of these particular tasks,
say about sentiment analysis,
about entailment and so on.
So typically for supervised learning,
we get a big corpus of such annotated or labeled data.
And then we feed that to a system
and the system is really trying to mimic,
so it's taking this input of the data
and then trying to mimic the output.
So it looks at an image and the human has tagged
that this image contains a banana
and now the system is basically trying to mimic that.
So that's its learning signal.
And so for supervised learning,
we try to gather lots of such data
and we train these machine learning models
to imitate the input output.
And the hope is basically by doing so,
now on unseen or like new kinds of data,
this model can automatically learn to predict these concepts.
So this is a standard sort of supervised setting.
For semi-supervised setting,
the idea typically is that you have,
of course, all of the supervised data,
but you have lots of other data which is unsupervised
or which is like not labeled.
Now the problem basically with supervised learning
and why you actually have all of these alternate
sort of learning paradigms is
supervised learning just does not scale.
So if you look at for computer vision,
the sort of largest one of the most popular datasets
is ImageNet, right?
So the entire ImageNet dataset
has about 22,000 concepts and about 14 million images.
So these concepts are basically just nouns
and they're annotated on images.
And this entire dataset was a mammoth data collection effort.
It actually gave rise to a lot of powerful learning algorithms
as credited with like sort of the rise
of deep learning as well.
But this dataset took about 22 human years
to collect, to annotate.
And it's not even that many concepts, right?
It's not even that many images.
14 million is nothing really.
Like you have about I think 400 million images or so
or even more than that uploaded to most of the popular
sort of social media websites today.
So now supervised learning just doesn't scale.
If I want to now annotate more concepts,
if I want to have this various types of fine-grained concepts,
then it won't really scale.
So now you come up to these sort of different learning paradigms,
for example, semi-supervised learning,
where the idea is, of course,
you have this annotated corpus of supervised data
and you have lots of these unlabeled images.
And the idea is that the algorithm should basically try
to measure some kind of consistency
or really try to measure some kind of signal
on this sort of unlabeled data
to make it self more confident
about what it's really trying to predict.
So by access to this lots of unlabeled data,
the idea is that the algorithm actually learns
to be more confident and actually gets better
at predicting these concepts.
And now we come to the other extreme,
which is like self-supervised learning.
The idea basically is that the machine
or the algorithm should really discover concepts
or discover things about the world
or learn representations about the world which are useful
without access to explicit human supervision.
So the word supervision is still in the term self-supervised.
So what is the supervision signal?
And maybe that perhaps is when Jan Lakoon and you argue
that unsupervised is the incorrect in terminology here.
So what is the supervision signal
when the humans aren't part of the picture
or not a big part of the picture?
Right.
So self-supervised, the reason it has the term
supervised in itself is because you're using the data
itself as supervision.
So because the data serves as its own source of supervision,
it's self-supervised in that way.
Now, the reason a lot of people,
I mean, we did it in that blog post with Jan,
but a lot of other people have also argued
for using this term self-supervised.
So starting from like 94,
from Virginia Desa's group, I think UCSD,
and now she's at UCSD.
Jitendra Malik has said this a bunch of times as well.
So you have supervised.
And then unsupervised basically means everything
which is not supervised.
But that includes stuff like semi-supervised,
that includes other like transductive learning,
lots of other sort of settings.
So that's the reason like now people are
preferring this term self-supervised
because it explicitly says what's happening.
The data itself is the source of supervision
and any sort of learning algorithm which tries to extract
just sort of data supervision signals from the data itself
is a self-supervised algorithm.
But there is within the data a set of tricks
which unlock the supervision.
So can you give me some examples?
And there's innovation, ingenuity required
to unlock that supervision.
The data doesn't just speak to you, some ground truth.
You have to do some kind of trick.
So I don't know what your favorite domain is.
So you specifically specialize in visual learning,
but is there favorite examples maybe in language
or other domains?
Perhaps the most successful applications
have been in NLP, not language processing.
So the idea basically being that you can train models
that you have a sentence and you mask out certain words.
And now these models learn to predict the masked out words.
So if you have the cat jumped over the dog,
so you can basically mask out cat.
And now you are essentially asking the model
to predict what was missing, what did I mask out?
So the model is going to predict basically
a distribution over all the possible words that it knows.
And probably it has like if it's a well-trained model,
it has a sort of higher probability density
for this word cat.
For vision, I would say the easier example, which
is not as widely used these days,
is basically say, for example, video prediction.
So video is, again, a sequence of things.
So you can ask the model, so if you
have a video of, say, 10 seconds,
you can feed in the first nine seconds to a model
and then ask it, hey, what happens basically in the 10
second?
Can you predict what's going to happen?
And the idea basically is because the model
is predicting something about the data itself.
Of course, you didn't need any human to tell you
what was happening because the 10 second video was naturally
captured because the model is predicting
what's happening there.
It's going to automatically learn something
about the structure of the world, how objects move,
object permanence, and these kinds of things.
So if I have something at the edge of the table,
it will fall down.
Things like these, which you really
don't have to sit and annotate.
In a supervised learning setting,
I would have to sit and annotate.
This is a cup.
Now I move this cup.
This is still a cup.
And now I move this cup.
It's still a cup.
And then it falls down.
And this is a fallen down cup.
So I won't have to annotate all of these things
in a self-supervised setting.
Isn't that kind of a brilliant little trick
of taking a series of data that is consistent
and removing one element in that series
and then teaching the algorithm to predict that element?
Isn't that, first of all, that's quite brilliant.
It seems to be applicable in anything
that has the constraint of being a sequence that
is consistent with the physical reality.
The question is, are there other tricks like this
that can generate the self-supervision signal?
So sequence is possibly the most widely used one in NLP.
For vision, the one that is actually
used for images, which is very popular these days,
is basically taking an image and now taking
different crops of that image.
So you can basically decide to crop, say, the top left corner
and you crop, say, the bottom right corner
and asking a network to basically present it
with a choice, saying that, OK, now you have this image.
You have this image.
Are these the same or not?
And so the idea, basically, is that because in an image,
different parts of the image are going to be related.
So for example, if you have a chair and a table,
basically these things are going to be closed by versus,
if you take, again, if you have a zoomed-in picture of a chair,
if you've taken different crops, it's
going to be different parts of the chair.
So the idea, basically, is that different crops
of the image are related.
And so the features or the representations
that you get from these different crops
should also be related.
So this is possibly the most widely
used trick these days for cell supervised
learning and computer vision.
So again, using the consistency that's
inherent to physical reality in visual domain,
that's parts of an image are consistent.
And then in the language domain or anything
that has sequences, like language or something
that's like a time series, then you
can chop off parts in time.
It's similar to the story of RNNs and CNNs,
of RNNs and covenants.
Yuwen Yan Lacoon wrote the blog post in March 2021 titled,
Self-Supervised Learning, the Dark Matter of Intelligence.
Can you summarize this blog post and maybe explain
the main idea or set of ideas?
The blog post was mainly about just telling,
I mean, this is really an accepted fact,
I would say, for a lot of people now
that self-supervised learning is something
that is going to be a plane important role for machine
learning algorithms that come in the future and even now.
Let me just comment that we don't yet
have a good understanding of what dark matter is.
That's true.
So the idea basically being.
The metaphor doesn't exactly transfer,
but maybe it's actually perfectly transfers
that we don't know.
We have an inkling that it'll be a big part
of whatever solving intelligence looks like.
Right.
So I think self-supervised learning, the way it's done
right now is, I would say, the first step towards what it
probably should end up learning or what it should enable us
to do.
So the idea for that particular piece
was self-supervised learning is going
to be a very powerful way to learn common sense about the world
or stuff that is really hard to label.
For example, is this piece over here heavier than the cup?
Now, for all these kinds of things,
you'll have to sit and label these things.
So supervised learning is clearly not going to scale.
So what is the thing that's actually going to scale?
It's probably going to be an agent that can either actually
interact with it to lift it up or observe me doing it.
So if I'm basically lifting these things up,
it can probably reason about, hey,
this is taking him more time to lift up
or the velocity is different.
Whereas the velocity for this is different,
probably this one is heavier.
So essentially by observations of the data,
you should be able to infer a lot of things about the world
without someone explicitly telling you, this is heavy.
This is not.
This is something that can pour.
This is something that cannot pour.
This is somewhere that you can sit.
This is not somewhere that you can sit.
But you just mentioned the ability
to interact with the world.
There's so many questions that are yet to be,
that are still open, which is, how do you
select a set of data over which the self-supervised learning
process works?
How much interactivity, like in the active learning
or the machine teaching context, is there,
what are the reward signals?
Like how much actual interaction there
is with the physical world, that kind of thing.
So that could be a huge question.
And then on top of that, which I have a million questions
about, which we don't know the answers to,
but it's worth talking about, is how much reasoning is
involved, how much accumulation of knowledge
versus something that's more akin to learning,
or whether that's the same thing.
So it is truly a dark matter.
We don't know how exactly to do it.
But we are, I mean, a lot of us are actually
convinced that it's going to be a sort of major thing
in machine learning.
So let me reframe it, then, that human supervision
cannot be at large scale, the source of the solution
to intelligence.
So the machines have to discover the supervision
in the natural signal of the world.
I mean, the other thing is also that humans are not
particularly good labors, they're not very consistent.
For example, what's the difference between a dining
table and a table?
Is it just the fact that one, if you just
look at a particular table, what makes us say one is dining
table and the other is not?
Humans are not particularly consistent,
they're not very good sources of supervision
for a lot of these kind of edge cases.
So it may be also the fact that if we want an algorithm
or want a machine to solve a particular task for us,
we can maybe just specify the end goal.
And the stuff in between, we really probably
should not be specifying because we're not maybe
going to confuse it a lot, actually.
Well, humans can't even answer the meaning of life.
So I'm not sure we're good supervisors of the end goal
either.
So let me ask you about categories.
Humans are not very good at telling the difference between what
is and isn't a table, like you mentioned.
Do you think it's possible, let me ask you like a pretend
you're Plato, is it possible to create a pretty good taxonomy
of objects in the world?
It seems like a lot of approaches in machine learning
kind of assume a hopeful vision that it's possible
to construct a perfect taxonomy.
Or it exists perhaps out of our reach,
but we can always get closer and closer to it.
Or is that a hopeless pursuit?
I think it's hopeless in some way.
So the thing is for any particular categorization
that you create, if you have a discrete sort of categorization,
I can always take the nearest two concepts
or I can take a third concept and I can blend it in
and I can create a new category.
So if you were to enumerate n categories,
I will always find an n plus 1 category for you.
That's not going to be in the n categories.
And I can actually create not just n plus 1,
I can very easily create far more than n categories.
The thing is a lot of things we talk about
are actually compositional.
So it's really hard for us to come and sit and enumerate
all of these out.
And they compose in various weird ways, right?
Like you have a croissant and a doughnut come together
to form a cronut.
So if you were to like enumerate all the foods up until,
I don't know, whenever the cronut was about 10 years ago
or 15 years ago, then this entire thing called
cronut would not exist.
Yeah, I remember there was the most awesome video
of a cat wearing a monkey costume.
Yeah, yes.
People should look it up, it's great.
So is that a monkey or is that a cat?
It's a very difficult philosophical question.
So there is a concept of similarity between objects.
So you think that can take us very far?
Just kind of getting a good function,
a good way to tell which parts of things are similar
and which parts of things are very different?
I think so, yeah.
So you don't necessarily need to name everything
or assign a name to everything to be able to use it, right?
So there are like lots of...
Shakespeare said that, what's in a name?
What's in a name, yeah, okay.
And I mean, lots of like, for example, animals, right?
They don't have necessarily a well-formed
like syntactic language,
but they're able to go about their day perfectly.
The same thing happens for us.
So, I mean, we probably look at things and we figure out,
oh, this is similar to something else that I've seen before.
And then I can probably learn how to use it.
So I haven't seen all the possible doorknobs in the world.
But if you show me, like I was able to get into
this particular place fairly easily,
I've never seen that particular doorknob.
So I, of course, related to all the doorknobs
that I've seen and I know exactly how it's going to open.
I have a pretty good idea of how it's going to open.
And I think this kind of translation between experiences
only happens because of similarity.
Because I'm able to relate it to a doorknob.
If I related it to a hairdryer,
I would probably be stuck still outside,
not able to get in.
Again, a bit of a philosophical question,
but can similarity take us all the way to understanding a thing?
Can having a good function that compares objects
get us to understand something profound
about singular objects?
I think I'll ask you a question back.
What does it mean to understand objects?
Well, let me tell you what that's similar to.
No.
So there's an idea of sort of reasoning
by analogy kind of thing.
I think understanding is the process of placing that thing
in some kind of network of knowledge that you have.
That it perhaps is fundamentally related to other concepts.
So it's not like understanding is fundamentally related
by composition of other concepts
and maybe in relation to other concepts.
And maybe deeper and deeper understanding
is maybe just adding more edges to that graph somehow.
So maybe it is a composition of similarities.
I mean, ultimately, I suppose it is a kind of embedding
in that wisdom space.
Yeah, okay, wisdom space is good.
I think I do think, right?
So similarity does get you very, very far.
Is it the answer to everything?
I mean, I don't even know what everything is,
but it's going to take us really far.
And I think the thing is things are similar
in very different contexts, right?
So an elephant is similar to, I don't know,
another sort of wild animal.
Let's just pick, I don't know, lion in a different way
because they're both four-legged creatures.
They're also land animals.
But of course, they're very different
in a lot of different ways.
So elephants are like herbivores, lions are not.
So similarity does, similarity and particularly dissimilarity
also sort of actually helps us understand a lot about things.
And so that's actually why I think
discrete categorization is very hard.
Just like forming this particular category of elephant
and a particular category of lion,
maybe it's good for like just like taxonomy,
biological taxonomies.
But when it comes to like other things
which are not as maybe, for example, like grilled cheese,
right? I have a grilled cheese I dip it in tomato
and I keep it outside.
Now, is that still a grilled cheese
or is that something else?
All right, so categorization is still very useful
for solving problems.
But is your intuition then sort of,
the self-supervised should be the,
to borrow Jan Lacoon's terminology,
should be the cake and then categorization,
the classification, maybe the supervised layer
should be just like the thing on top,
the cherry or the icing or whatever.
So if you make it the cake, it gets in the way of learning.
If you make it the cake, then you don't,
we won't be able to sit and annotate everything.
That's as simple as it is.
Like that's my very practical view on it.
It's just, I mean, in my PhD,
I sat down and annotated like a bunch of cars
for one of my projects.
And very quickly I was just like,
it was in a video and I was basically drawing boxes
around all these cars.
And I think I spent about a week doing all of that
and I barely got anything done.
And basically this was, I think my first year of my PhD
at like a second year of my master's.
And then by the end of it, I'm like, okay,
this is just hopeless.
I can keep doing it.
And when I'd done that, someone came up to me
and they basically told me,
oh, this is a pickup truck.
This is not a car.
And that's like, aha, this actually makes sense
because a pickup truck is not really like,
what was I annotating?
Was I annotating anything that is mobile?
Or was I annotating particular sedans
or was I annotating SUVs?
What was I doing?
By the way, the annotation was bounding boxes?
Bounding boxes.
There's so many deep, profound questions here.
You're almost cheating your way out of it
by doing self-supervised learning, by the way,
which is like, what makes for an object?
As opposed to solve intelligence,
maybe you don't ever need to answer that question.
I mean, this is the question that anyone
that's ever done an annotation because it's so painful
gets to ask like, why am I doing a drawing,
very careful line around this object?
Like what is the value?
I remember when I first saw semantic segmentation
where you have like instant segmentation
where you have a very exact line
around the object in a 2D plane
of a fundamentally 3D object projected on a 2D plane.
So you're drawing a line around a car
that might be occluded,
there might be another thing in front of it,
but you're still drawing the line
of the part of the car that you see.
How is that the car?
Why is that the car?
Like I had like an existential crisis every time.
Like how is that going to help us understand
a solve computer vision?
I'm not sure I have a good answer to what's better.
And I'm not sure I share the confidence that you have
that self-supervised learning can take us far.
I think I'm more and more convinced
that it's a very important component,
but I still feel like we need to understand what makes,
like this dream of maybe what it's called symbolic AI
of arriving, like once you have this common sense base,
be able to play with these concepts
and build graphs or hierarchies of concepts on top,
in order to then form a deep sense
of this three-dimensional world or four-dimensional world
and be able to reason and then project that
onto 2D playing in order to interpret a 2D image.
Can I ask you just an out there question?
I remember, I think Andre Capati had a blog post
about computer vision, like being really hard.
I forgot what the title was, but it's many, many years ago.
And he had, I think President Obama stepping on a scale
and there was humor and there was a bunch of people
laughing and whatever.
And there's a lot of interesting things about that image
and I think Andre highlighted a bunch of things
about the image that us humans are able to immediately
understand, like the idea, I think of gravity
and that you have the concept of a weight.
You immediately project, because of our knowledge of pose
and how human bodies are constructed,
you understand how the forces are being applied
with the human body.
They're really interesting.
Other thing that you're able to understand
is multiple people looking at each other in the image.
You're able to have a mental model
of what the people are thinking about.
You're able to infer like, oh, this person is probably
thinks like is laughing at how humorous the situation is.
And this person is confused about what the situation is
because they're looking this way.
We're able to infer all of that.
So that's human vision.
How difficult is computer vision?
Like in order to achieve that level of understanding
and maybe how big of a part
does self-supervised learning play in that, do you think?
And do you still, back that was like over a decade ago,
I think Andre and I think a lot of people agreed
that computer vision is really hard.
Do you still think computer vision is really hard?
I think it is, yes.
And getting to that kind of understanding,
I mean, it's really out there.
So if you ask me to solve just that particular problem,
I can do it the supervised learning route.
I can always construct a data set and basically predict,
oh, is there humor in this or not?
And of course I can do it.
Actually, that's a good question.
Do you think you can, okay, okay.
Do you think you can do human supervised annotation
of humor?
To some extent, yes.
I'm sure it'll work.
I mean, it won't be as bad as like randomly guessing.
I'm sure it can still predict whether it's humorous
or not in some way.
Yeah, maybe like Reddit upvotes is the signal.
I don't know.
I mean, it won't do a great job, but it'll do something.
It may actually be like it may find certain things
which are not humorous, humorous as well,
which is going to be bad for us.
But I mean, it'll do a, it won't be random.
Yeah, kind of like my sense of humor.
Okay, so fine.
So you can, that particular problem, yes.
But the general problem you're saying is hard.
The general problem is hard.
And I mean, self-supervised learning
is not the answer to everything.
Of course it's not.
I think if you have machines that are going to communicate
with humans at the end of it,
you want to understand what the algorithm is doing, right?
You want it to be able to like produce an output
that you can decipher, that you can understand,
or it's actually useful for something else,
which again is a human.
So at some point in this sort of entire loop,
a human steps in.
And now this human needs to understand what's going on.
And at that point, this entire notion of language
or semantics really comes in.
If the machine just spits out something,
and if we can't understand it,
then it's not really that useful for us.
So self-supervised learning is probably going to be useful
for a lot of the things before that part,
before the machine really needs to communicate
a particular kind of output with a human.
Because, I mean, otherwise,
how is it going to do that without language?
Or some kind of communication.
But you're saying that it's possible to build
a big base of understanding or whatever of,
what's it about?
Concepts.
Concepts, yeah.
Like common sense concepts.
Supervised learning in the context of computer vision
is something you've focused on,
but that's a really hard domain.
And it's kind of the cutting edge
of what we're as a community working on today.
Can we take a little bit of a step back
and look at language?
Can you summarize the history of success
of self-supervised learning in natural language processing,
language modeling?
What are transformers?
What is the masking, the sentence completion
that you mentioned before?
How does it lead us to understand anything?
Semantic meaning of words,
syntactic role of words and sentences.
So I'm, of course, not the expert in NLP.
I kind of follow it a little bit from the sides.
So the main sort of reason
why all of this masking stuff works is,
I think it's called the distributional hypothesis in NLP.
The idea basically being that words
that occur in the same context should have similar meaning.
So if you have the blank jumped over the blank,
it basically, whatever is like in the first blank
is basically an object that can actually jump
is going to be something that can jump.
So a cat or a dog or I don't know, sheep, something,
all of these things can basically be
in that particular context.
And now, so essentially the idea is that
if you have words that are in the same context
and you predict them, you're going to learn
a lot of useful things about how words are related
because you're predicting by looking at their context
what the word is going to be.
So in this particular case, the blank jumped over the fence.
So now if it's a sheep, the sheep jumped over the fence,
the dog jumped over the fence.
So essentially the algorithm or the representation
basically puts together these two concepts together.
So it says, okay, dogs are going to be kind of related to sheep
because both of them occur in the same context.
Of course, now you can decide depending
on their particular application downstream,
you can say that dogs are absolutely not related to sheep
because well, I don't, I really care about dog food,
for example, I'm a dog food person
and I really want to give this dog food
to this particular animal.
So depending on what your downstream application is,
of course, this notion of similarity or this notion
or this common sense that you've learned
may not be applicable.
But the point is basically that this,
just predicting what the blanks are
is going to take you really, really far.
So there's a nice feature of language
that the number of words in a particular language
is very large, but it's finite
and it's actually not that large
in the grand scheme of things.
I still got up because we take it for granted.
So first of all, when you say masking,
you're talking about this very process
of the blank of removing words from a sentence
and then having the knowledge of what word went there
in the initial data set.
That's the ground truth that you're training on
and then you're asking the neural network
to predict where it goes there.
That's like a little trick.
It's a really powerful trick.
The question is how far that takes us
and the other question is, is there other tricks?
Because to me, it's very possible
there's other very fascinating tricks.
I'll give you an example in autonomous driving,
there's a bunch of tricks
that give you the self-supervised signal back.
For example, very similar to sentences,
but not really, which is you have signals
from humans driving the car
because a lot of us drive cars to places.
And so you can ask the neural network to predict
what's going to happen the next two seconds
for a safe navigation through the environment.
And the signal comes from the fact
that you also have knowledge of what happened
in the next two seconds because you have video of the data.
The question in autonomous driving, as it is in language,
can we learn how to drive autonomously
based on that kind of self-supervision?
Probably the answer is no.
The question is how good can we get?
And the same with language, how good can we get?
And are there other tricks?
Like we get sometimes super excited
by this trick that works really well,
but I wonder, it's almost like mining for gold.
I wonder how many signals there are in the data
that could be leveraged that are like there.
Is that, I just want to kind of linger on that
because sometimes it's easy to think
that maybe this masking process is self-supervised learning.
No, it's only one method.
So there could be many, many other methods,
many tricky methods, maybe interesting ways
to leverage human computation in very interesting ways
that might actually border on semi-supervised learning,
something like that.
Obviously, the internet is generated by humans
at the end of the day.
So all that to say is what's your sense
in this particular context of language,
how far can that masking process take us?
So it has stood the test of time, right?
I mean, so word-to-vec, the initial sort of NLP technique
that was using this to now, for example,
like all the BERT and all these big models
that we get, BERT and Roberta, for example,
all of them are still sort of based
on the same principle of masking.
It's taken us really far.
I mean, you can actually do things like,
oh, these two sentences are similar or not,
whether this particular sentence follows this other sentence
in terms of logic or entailment.
You can do a lot of these things with this,
just this masking trick.
So I'm not sure if I can predict how far it can take us
because when it first came out, when word-to-vec was out,
I don't think a lot of us would have imagined
that this would actually help us do some kind
of entailment problems and really that well.
And so just the fact that by just scaling up
the amount of data that we're training on
and using better and more powerful neural network
architectures has taken us from that to this,
is just showing you how maybe poor predictors we are.
As humans, how poor we are at predicting
how successful a particular technique is going to be.
So I think I can say something now,
but like 10 years from now,
I look completely stupid basically predicting this.
In the language domain, is there something in your work
that you find useful and insightful
and transferable to computer vision,
but also just, I don't know, beautiful and profound
that I think carries through to the vision domain?
I mean, the idea of masking has been very powerful.
It has been used in vision as well for predicting,
like you say, the next sort of,
if you have in sort of frames,
then you predict what's going to happen in the next frame.
So that's been very powerful.
In terms of modeling, like in just terms
in terms of architecture,
I think you would have asked about transformers
a while back.
That has really become,
like it has become super exciting for computer vision now.
Like in the past, I would say year and a half,
it's become really powerful.
What's a transformer?
Right.
I mean, the core part of a transformer
is something called the self-attention model.
So it came out of Google.
And the idea basically is that if you have n elements,
what you're creating is a way
for all of these n elements to talk to each other.
So the idea basically is that you are paying attention.
Each element is paying attention
to each of the other element.
And basically by doing this,
it's really trying to figure out,
you're basically getting a much better view of the data.
So for example, if you have a sentence of like four words,
the point is if you get a representation
or a feature for this entire sentence,
it's constructed in a way
such that each word has paid attention to everything else.
Now, the reason it's like different from say,
what you would do in a ConvNet
is basically that in the ConvNet,
you would only pay attention to a local window.
So each word would only pay attention to its next neighbor
or like one neighbor after that.
And the same thing goes for images.
In images, you would basically pay attention to pixels
in a three cross three or a seven cross seven neighborhood.
And that's it.
Whereas with a transformer, the self-attention mainly,
the sort of idea is that each element needs
to pay attention to each other element.
And when you say attention,
maybe another way to phrase that
is you're considering a context,
a wide context in terms of the wide context of the sentence
in understanding the meaning of a particular word
and a computer vision that's understanding a larger context
to understand the local pattern
of a particular local part of an image.
Right.
So basically if you have say,
again a banana in the image,
you're looking at the full image first.
So whether it's like, you know,
you're looking at all the pixels that are off a kitchen,
off a dining table and so on.
And then you're basically looking at the banana also.
Yeah.
By the way, in terms of if we were to train
the funny classifier,
there's something funny about the word banana.
Just going to anticipate that my-
I am wearing a banana shirt, so yeah.
Oh, is there bananas on it?
Okay, so masking has worked for the vision context as well.
And so this transformer idea has worked as well.
So basically looking at all the elements
to understand a particular element
has been really powerful in vision.
The reason is like a lot of things
when you're looking at them in isolation.
So if you look at just a blob of pixels,
so Antonio Turalba at MIT used to have this
like really famous image,
which I looked at when I was a PhD student,
where he would basically have a blob of pixels
and he would ask you, hey, what is this?
And it looked basically like a shoe
or like it could look like a TV remote.
It could look like anything.
And it turns out it was a beer bottle.
But I'm not sure.
It was one of these three things,
but basically he showed you the full picture
and then it was very obvious what it was.
But the point is just by looking at
that particular local window, you couldn't figure out.
Because of resolution, because of other things,
it's just not easy always to just figure out
by looking at just the neighborhood of pixels,
what these pixels are.
And the same thing happens for language as well.
For the parameters that have to learn
something about the data,
you need to give it the capacity
to learn the essential things.
Like if it's not actually able to receive the signal at all,
then it's not gonna be able to learn that signal.
And to understand images, to understand language,
you have to be able to see words in their full context.
Okay, what is harder to solve, vision or language?
Visual intelligence or linguistic intelligence?
So I'm going to say computer vision is harder.
My reason for this is basically that
language of course has a big structure to it
because we developed it.
Whereas vision is something that is common
in a lot of animals.
Everyone is able to get by,
a lot of these animals on Earth
are actually able to get by without language.
And a lot of these animals,
we also deem to be intelligent.
So clearly intelligence does have
like a visual component to it.
And yes, of course in the case of humans,
it of course also has a linguistic component.
But it means that there is something far more fundamental
about vision than there is about language.
And I'm sorry to anyone who disagrees,
but yes, this is what I feel.
So that's being a little bit reflected
in the challenges that have to do with the progress
of self supervised learning, would you say?
Or is that just the peculiar accidents
of the progress of the AI community
that we focused on or we discovered self-attention
and transformers in the context of language first?
So like the self supervised learning success was actually,
for vision has not much to do with the transformers part.
I would say it's actually been independent a little bit.
I think it's just that the signal was a little bit different
for vision than there was for like NLP
and probably NLP folks discovered it before.
So for vision, the main success has basically been
just like crops so far, like taking different crops
of images, whereas for NLP, it was this masking thing.
But also the level of success is still much higher
for language.
Yes, it has.
So that has a lot to do with,
I mean, I can get into a lot of details
for this particular question, let's go for it.
Okay, so the first thing is language is very structured.
So you are going to produce a distribution
over a finite vocabulary.
English has a finite number of words.
It's actually not that large.
And you need to produce basically,
when you're doing this masking thing,
all you need to do is basically tell me
which one of these like 50,000 words it is.
That's it.
Now for vision, let's imagine doing the same thing.
Okay, we're basically going to blank out
a particular part of the image.
And we ask the network or this neural network
to predict what is present in this missing patch.
It's combinatorially large, right?
You have 256 pixel values.
If you're even producing basically a seven cross seven
or a 14 cross 14 like window of pixels,
at each of these 169 or each of these 49 locations,
you have 256 values to predict.
And so it's really, really large.
And very quickly, the kind of like prediction problems
that we're setting up are going to be extremely
like intractable for us.
And so the thing is for NLP, it has been really successful
because we are very good at predicting,
like doing this like distribution over a finite set.
And the problem is when this set becomes really large,
we are going to become really, really bad
at making these predictions.
And at solving basically this particular set of problems.
So if you were to do it exactly in the same way as NLP
for vision, there is very limited success.
The way stuff is working right now
is actually not by predicting these masks.
It's basically by saying that you take these two
like crops from the image,
you get a feature representation from it.
And just saying that these two features,
so they're like vectors,
just saying that the distance between these vectors
should be small.
And so it's a very different way of learning
from the visual signal than there is from NLP.
Okay, the other reason is the distributional hypothesis
that we talked about for NLP, right?
So a word given its context,
basically the context actually supplies
a lot of meaning to the word.
Now, because there are just finite number of words
and there is a finite way in which we compose them,
of course, the same thing holds for pixels,
but in language there's a lot of structure, right?
So I always say whatever,
the dash jumped over the fence, for example.
There are lots of these sentences that you'll get.
And from this, you can actually look
at this particular sentence might occur
in a lot of different contexts as well.
This exact same sentence might occur in a different context.
So the sheep jumped over the fence,
the cat jumped over the fence,
the dog jumped over the fence.
So you immediately get a lot of these words,
which are, because this particular token itself
has so much meaning, you get a lot of these tokens
or these words which are actually going to have
this related meaning across, given this context.
Whereas for vision, it's much harder.
Because just by pure, the way we capture images,
lighting can be different.
There might be different noise in the sensor.
So the thing is you're capturing a physical phenomenon,
and then you're basically going through
a very complicated pipeline of image processing,
and then you're translating that into
some kind of digital signal.
Whereas with language, you write it down,
and you transfer it to a digital signal,
almost like it's a lossless transfer.
And each of these tokens are very, very well-defined.
There could be a little bit of an argument there,
because language, as written down, is a projection of thought.
This is one of the open questions,
is if you perfectly can solve language,
are you getting close to being able to solve,
easily with flying colors past the Turing test
kind of thing.
So that's, it's similar, but different
and the computer vision problem is in the 2D plane,
is a projection to be to mention a world.
So perhaps there are similar problems there.
Maybe there's a good, yeah.
I think what I'm saying is NLP is not easy,
of course, don't get me wrong.
Abstract thought expressed in knowledge,
or knowledge basically expressed in language,
is really hard to understand, right?
I mean, we've been communicating with language for so long,
and it is, of course, a very complicated concept.
The thing is, at least getting some more reasonable,
like being able to solve some kind of reasonable tasks
with language, I would say slightly easier
than it is with computer vision.
Yeah, I would say, yeah.
So that's well put.
I would say getting impressive performance on language
is easier.
I feel like for both language and computer vision,
there's going to be this wall of like,
like this hump you have to overcome
to achieve super human level performance,
or human level performance.
And I feel like for language, that wall is farther away.
So you can get pretty nice,
you can do a lot of tricks.
You can show really impressive performance.
You can even fool people that you're tweeting,
or you're blog posts writing,
or your question answering has intelligence behind it.
But to truly demonstrate understanding of dialogue,
of continuous long form dialogue,
that would require perhaps big breakthroughs.
In the same way in computer vision,
I think the big breakthroughs need to happen earlier
to achieve impressive performance.
This might be a good place to, you already mentioned it,
but what is contrastive learning,
and what are energy-based models?
Contrastive learning is sort of a paradigm of learning
where the idea is that you are learning
this embedding space,
or so you're learning this sort of vector space
of all your concepts.
And the way you learn that is basically by contrasting.
So the idea is that you have a sample.
You have another sample that's related to it.
So that's called the positive,
and you have another sample that's not related to it.
So that's negative.
So for example, let's just take an NLP,
and a simple example in computer vision.
So you have an image of a cat,
you have an image of a dog,
and for whatever application that you're doing,
say you're trying to figure out what pets are,
you think that these two images are related.
So image of a cat and dog are related,
but now you have another third image of a banana,
because you don't like that word.
So now you basically have this banana.
Thank you for speaking to the crowd.
And so you take both of these images,
and you take the image from the cat,
the image from the dog,
you get a feature from both of them.
And now what you're training the network to do
is basically pull both of these features together
while pushing them away from the feature of a banana.
So this is the contrastive part.
So you're contrasting against the banana.
So there's always this notion of a negative and a positive.
Now, energy-based models are like one way
that Jan sort of explains a lot of these methods.
So Jan basically, I think a couple of years or more than that,
like when I joined Facebook,
Jan used to keep mentioning this word energy-based models.
And of course, I had no idea what he was talking about.
So then one day I caught him in one of the conference rooms
and I'm like, can you please tell me what this is?
So then like very patiently,
he sat down with like a marker and a whiteboard.
And his idea basically is that
rather than talking about probability distributions,
you can talk about energies of models.
So models are trying to minimize certain energies
in certain space,
or they're trying to maximize a certain kind of energy.
And the idea basically is that
you can explain a lot of the contrastive models,
GANs, for example, which are like
Generative Adversarial Networks.
A lot of these modern learning methods,
or VIEs, which are Variational Autoencoders,
you can really explain them very nicely
in terms of an energy function
that they're trying to minimize or maximize.
And so by putting this common sort of language
for all of these models,
what looks very different in machine learning
that oh, VIEs are very different from what GANs are,
are very different from what contrastive models are,
you actually get a sense of like,
oh, these are actually very, very related.
It's just that the VIE or the mechanism
in which they're sort of maximizing
or minimizing this energy function is slightly different.
It's revealing the commonalities between all these approaches
and putting a sexy word on top of it, like energy.
And so similarities,
two things that are similar have low energy,
like the low energy signifying similarity.
Right, exactly.
So basically the idea is that if you were to imagine
like the embedding as a manifold, a 2D manifold,
you would get a hill or like a high sort of peak
in the energy manifold,
wherever two things are not related.
And basically you would have like a dip
where two things are related.
So you'd get a dip in the manner.
And in the self-supervised context,
how do you know two things are related
and two things are not related?
Right, so this is where all the sort of ingenuity
or tricks comes in, right?
So for example, like you can take the fill-in-the-blank problem
or you can take in the context problem.
And what you can say is two words
that are in the same context are related.
Two words that are in different contexts are not related.
For images, basically two crops from the same image
are related and whereas a third image is not related at all.
For a video, it can be two frames from that video
are related because they're likely to contain
the same sort of concepts in them,
whereas a third frame from a different video is not related.
So it basically is, it's a very general term.
Contrast-reducing is nothing really
to do with self-supervised learning.
It actually is very popular in, for example,
like any kind of metric learning
or any kind of embedding learning.
So it's also used in supervised learning.
It's also, and the thing is,
because we are not really using labels
to get these positive or negative pairs,
it can basically also be used for self-supervised learning.
So you mentioned one of the ideas in the vision context
that works is to have different crops.
So you could think of that as a way
to sort of manipulating the data to generate examples
that are similar.
Obviously, there's a bunch of other techniques.
You mentioned lighting as a very,
in images, lighting is something that varies a lot
and you can artificially change those kinds of things.
There's the whole broad field of data augmentation
which manipulates images in order to increase arbitrarily
the size of the dataset.
First of all, what is data augmentation?
And second of all, what's the role of data augmentation
in self-supervised learning and contrastive learning?
So data augmentation is just a way, like you said,
it's basically a way to augment the data.
So you have, say, end samples
and what you do is you basically define
some kind of transforms for the sample.
So you take your, say, image
and then you define a transform
where you can just increase, say, the colors,
like the colors or the brightness of the image
or increase or decrease the contrast of the image,
for example, or take different crops of it.
So data augmentation is just a process
to basically perturb the data or augment the data.
And so it has played a fundamental role
for computer vision for self-supervised learning,
especially the way most of the current methods
were contrastive or otherwise is by taking an image,
in the case of images, is by taking an image
and then computing basically two perturbations of it.
So these can be two different crops of the image
with different types of lighting
or different contrasts or different colors.
So you jitter the colors a little bit and so on.
And now the idea is basically
because it's the same object
or because it's like related concepts
in both of these perturbations,
you want the features from both of these perturbations
to be similar.
So now you can use a variety of different ways
to enforce this constraint,
like these features being similar.
You can do this by contrastive learning.
So basically both of these things are positives,
a third sort of image is negative.
You can do this basically by like clustering.
For example, you can say that both of these images
should, the features from both of these images
should belong in the same cluster because they're related.
Whereas image, like another image
should belong to a different cluster.
So there's a variety of different ways
to basically enforce this particular constraint.
By the way, when you say features,
it means there's a very large neural network
that extracting patterns from the image
and the kind of patterns that extracts
should be either identical or very similar.
That's what that means.
So the neural network basically takes in the image
and then outputs a set of like basically a vector
of like numbers and that's the feature.
And you want this feature for both of these
like different crops that you computed to be similar.
So you want this vector to be identical
in its like entries, for example.
Be like literally close in this multi-dimensional space
to each other.
Right.
And like you said, close can mean
a part of the same cluster or something like that
in this large space.
First of all that, I wonder if there is connection
to the way humans learn to this.
Almost like maybe subconsciously.
In order to understand a thing,
you kind of have to see it from two, three multiple angles.
I wonder, there's a lot of friends who are neuroscientists
maybe or incognito scientists.
I wonder if that's in there somewhere.
Like in order for us to place a concept
in its proper place, we have to basically crop it
in all kinds of ways, do basic data augmentation on it
in whatever very clever ways that the brain likes to do.
Right.
Like spinning around in our mind somehow
that that is very effective.
So I think for some of them, we like need to do it.
So like babies, for example, pick up objects,
like move them, put them, go to Dira and whatnot.
But for certain other things,
actually we are good at imagining it as well, right?
So if you, I have never seen, for example,
an elephant from the top.
I've never basically looked at it from top down.
But if you showed me a picture of it,
I could very well tell you that that's an elephant.
So I think some of it, we just like,
we naturally build it or transfer it
from other objects that we've seen
to imagine what it's going to look like.
Has anyone done that with the augmentation?
Like imagine all the possible things
that are occluded or not there,
but not just like normal things, like wild things,
but that are nevertheless physically consistent.
So, I mean, people do kind of like
occlusion based augmentation as well.
So you place in like a random box, gray box
to sort of mask out a certain part of the image.
And the thing is basically you're kind of occluding it.
For example, you place it say on half of a person's face.
So basically saying that, you know,
something below their nose is occluded,
because it's grayed out.
So.
No, I meant like, you have like, what is it?
A table and you can't see behind the table.
And you imagine there's a bunch of elves
with bananas behind the table.
Like, I wonder if there's useful to have
a wild imagination for the network,
because that's possible, well, maybe not elves,
but like puppies and kittens or something like that.
Just have a wild imagination
and like constantly be generating that wild imagination.
Cause in terms of data augmentation
that's currently applied, it's super ultra, very boring.
It's very basic data augmentation.
I wonder if there's a benefit to being wildly imaginable
while trying to be consistent with physical reality.
I think it's a kind of a chicken and egg problem, right?
Because to have like amazing data augmentation,
you need to understand what the scene is.
And what we're trying to do data augmentation
to learn what a scene is anyway.
So it's basically just keeps going on.
Before you understand it, just put elves with bananas
until you know it's not to be true.
Just like children have a wild imagination
until the adults ruin it all.
Okay, so what are the different kinds of data augmentation
that you've seen to be effective in visual intelligence?
For like vision, it's a lot of these image filtering operations.
So like blurring the image,
you know, all the kind of Instagram filters
that you can think of.
So like arbitrarily like make the red super red,
make the green super greens, like saturate the image.
Rotation cropping.
Rotation cropping, exactly.
All of these kind of things.
Like I said, lighting is a really interesting one to me.
Like that feels like really complicated to do.
So I mean, the augmentations that we work on
aren't like that involved.
So they're not going to be like physically realistic versions
of lighting.
It's not that you're assuming that there's a light source up
and then you're moving it to the right
and then what does the thing look like?
It's really more about like brightness of the image,
overall brightness of the image
or overall contrast of the image and so on.
But this is a really important point to me.
I always thought that data augmentation
holds an important key to big improvements
in machine learning.
And it seems that it is an important aspect
of self-supervised learning.
So I wonder if there's big improvements to be achieved
on much more intelligent kinds of data augmentation.
For example, currently,
maybe you can correct me if I'm wrong,
data augmentation is not parametrized.
Yeah.
You're not learning.
To me, it seems like data augmentation potentially
should involve more learning
than the learning process itself.
Right.
You're almost like thinking of like generative kind of,
it's the elves with bananas.
You're trying to, it's like very active imagination
of messing with the world
and teaching that mechanism for messing with the world
to be realistic.
Right.
Because that feels like,
I mean, it's imagination.
Just as you said, it feels like us humans are able to,
maybe sometimes subconsciously imagine,
before we see the thing,
imagine what we're expecting to see.
Like maybe several options.
And especially, we probably forgot,
but when we're younger,
probably the possibilities were wild,
they're more numerous.
And then as we get older,
we become to understand the world
and the possibilities of what we might see
becomes less and less and less.
So I wonder if you think there's a lot of breakthroughs
yet to be had in data augmentation.
And maybe also can you just comment on the stuff we have?
Is that a big part of self-supervised learning?
Yes, so data augmentation is like
key to self-supervised learning.
That has like the kind of augmentation that we're using.
And basically the fact that we're trying to learn
these neural networks that are predicting these features
from images that are robust under data augmentation
has been the key for visual self-supervised learning.
And they play a fairly fundamental role to it.
Now, the irony of all of this
is that deep learning purists will say,
the entire point of deep learning
is that you feed in the pixels to the network,
neural network, and it should figure out
the patterns on its own.
So if it really wants to look at edges,
it should look at edges.
You shouldn't really go and handcraft these features, right?
You shouldn't go tell it that look at edges.
So data augmentation should basically be
in the same category, right?
Why should we tell the network
or tell this entire learning paradigm?
What kinds of data augmentation that we're looking for?
We are encoding a very sort of human specific bias there.
That we know things are like,
if you change the contrast of the image,
it should still be an apple
or it should still see apple, not banana.
And basically if we change like colors,
it should still be the same kind of concept.
Of course, this is not one,
this doesn't feel like super satisfactory
because a lot of our human knowledge
or our human supervision
is actually going into the data augmentation.
So although we're calling it self-supervised learning,
a lot of the human knowledge
is actually being encoded in the data augmentation process.
So it's really like we've kind of sneaked away
the supervision at the input
and we're like really designing these nice list
of data augmentations that are working very well.
Of course, the idea is that it's much easier
to design a list of data augmentation that it is.
So humans are doing, never the less,
doing less and less work
and maybe leveraging their creativity more and more.
When we say data augmentation is not parametrized,
it means it's not part of the learning process.
Do you think it's possible to integrate
some of the data augmentation into the learning process?
I think so, I think so.
And in fact, it will be really beneficial for us
because a lot of these data augmentation
that we use in vision are very extreme.
For example, like when you have certain concepts,
again a banana, you take the banana
and then basically you change the color of the banana, right?
So you make it a purple banana.
Now this data augmentation process
is actually independent of the,
it has no notion of what is present in the image.
So it can change this color arbitrarily.
It can make it a red banana as well.
And now what we're doing is we're telling the neural network
that this red banana and,
so crop of this image which has the red banana
and a crop of this image where I change the color
to a purple banana should be,
the features should be the same.
Now bananas aren't red or purple, mostly.
So really the data augmentation process
should take into account what is present in the image
and what are the kinds of physical realities that are possible.
It shouldn't be completely independent of the image.
So you might get big gains if you,
instead of being drastic,
do subtle augmentation but realistic augmentation.
Right, realistic.
I'm not sure if it's subtle but like realistic for sure.
If it's realistic then even subtle augmentation
will give you big benefits.
Exactly, yeah.
And it will be like for particular domains
you might actually see like,
if for example now we're doing medical imaging,
there are going to be certain kinds of like geometric augmentation
which are not really going to be very valid for the human body.
So if you were to like actually loop in data augmentation
into the learning process,
it will actually be much more useful.
Now this actually does take us to maybe a semi-supervised
kind of a setting because you do want to understand
what is it that you're trying to solve.
So currently self-supervised learning
kind of operates in the wild, right?
So you do the self-supervised learning
and the purists and all of us basically say that,
okay, this should learn useful representations
and they should be useful for any kind of end task,
no matter it's like banana recognition
or like autonomous driving.
Now it's a tall order.
Maybe the first baby step for us should be that,
okay, if you're trying to loop in this data augmentation
into the learning process,
then we at least need to have some sense
of what we're trying to do.
Are we trying to distinguish between different types
of bananas or are we trying to distinguish
between banana and apple
or are we trying to do all of these things at once?
And so some notion of like what happens at the end
might actually help us do much better at this side.
Let me ask you a ridiculous question.
If I were to give you like a black box,
like a choice to have an arbitrary large data set
of real natural data
versus really good data augmentation algorithms,
which would you like to train in a self-supervised way on?
So natural data from the internet are arbitrary large.
So unlimited data.
Or it's like more controlled good data augmentation
on the finite data set.
The thing is like because our learning algorithms
for vision right now really rely on data augmentation,
even if you were to give me like an infinite source
of like image data,
I still need a good data augmentation algorithm.
You need something that tells you
that two things are similar.
Right.
And so something because you've given me
an arbitrarily large data set,
I still need to use data augmentation
to take that image construct like these two perturbations
of it and then learn from it.
So the thing is our learning paradigm
is very primitive right now.
Even if you were to give me lots of images,
it's still not really useful.
A good data augmentation algorithm
is actually going to be more useful.
So you can like reduce down the amount of data
that you'll give me by like 10 times.
But if you were to give me a good data augmentation
algorithm that will probably do better
than giving me like 10 times the size of that data,
but me having to rely on like a very primitive
data augmentation algorithm.
Like through tagging and all those kinds of things,
is there a way to discover things
that are semantically similar on the internet?
Obviously there is, but it might be extremely noisy.
And the difference might be farther away
than you would be comfortable with.
So I mean, yes, tagging will help you a lot.
It'll actually go a very long way
in figuring out what images are related or not.
And then so, but then the purists would argue
that when you're using human tags
because these tags are like supervision,
is it really self supervised learning now?
Because you're using human tags to figure out
which images are like similar.
Hashtag no filter means a lot of things.
Yes, I mean, there are certain tags
which are going to be applicable pretty much to anything.
So they're pretty useless for learning.
But I mean, certain tags are actually like
DI filter, for example, or the Taj Mahal, for example.
These tags are like very indicative of what's going on.
And they are, I mean, they are human supervision.
Yeah, this is one of the tasks of discovering
from human generated data strong signals
that could be leveraged for self supervision.
Like humans are doing so much work already.
Like many years ago, there was something that was called
I guess human computation back in the day.
Humans are doing so much work.
It'd be exciting to discover ways to leverage
the work they're doing to teach machines
without any extra effort from them.
An example could be like we said,
the driving, human's driving and machines can learn
from the driving.
I always hoped that there could be some supervision signal
discovered in video games because there's so many people
that play video games that it feels like so much effort
is put into video games, into playing video games.
And you can design video games somewhat cheaply
to include whatever signals you want.
It feels like that could be leveraged somehow.
So people are using that, like there are actually folks
right here in UT Austin, like Phillip Cranbull
is a professor at UT Austin.
He's been like working on video games
as a source of supervision.
I mean, it's really fun, like as a PhD student
getting to basically play video games all day.
Yeah, but so I do hope that kind of thing scales
and like ultimately boils down to discovering
some undeniably very good signal.
It's like masking in an LP.
But that said, there's non-contrastive methods.
Right.
What do non-contrastive, energy-based, self-supervised
learning methods look like and why are they promising?
So like I said about contrastive learning,
you have this notion of a positive and a negative.
Now, the thing is this entire learning paradigm
really requires access to a lot of negatives
to learn a good sort of feature space.
The idea is if I tell you, OK, so a cat and a dog are similar
and they're very different from a banana.
The thing is this is a fairly simple analogy, right?
Because, well, bananas look visually very different
from what cats and dogs do.
So very quickly, if this is the only source of supervision
that I'm giving you, your learning is not going to be
like after a point, then neural network is really
not going to learn a lot because the negative that you're
getting is going to be so random.
So it can be, oh, a cat and a dog are similar,
but they're very different from a Volkswagen Beetle.
Now, this car looks very different from these animals
again.
So the thing is in contrastive learning,
the quality of the negative sample really matters a lot.
And so what has happened is basically
that typically these methods that are contrastive
really require access to lots of negatives, which
becomes harder and harder to sort of scale
when designing a learning algorithm.
So that's been one of the reasons
why non-contrastive methods have become popular
and why people think that they're going to be more useful.
So a non-contrastive method, for example,
like clustering is one non-contrastive method,
the idea basically being that you have two of these samples.
So the cat and dog or two crops of this image,
they belong to the same cluster.
And so essentially, you're basically
doing clustering online when you're learning this network
and which is very different from having access
to a lot of negatives explicitly.
The other way which has become really popular
is something called self-distillation.
So the idea basically is that you have a teacher network
and a student network.
And the teacher network produces a feature.
So it takes in the image.
And basically, the neural network
figures out the patterns, gets the feature out.
And there's another neural network
which is the student neural network,
and that also produces a feature.
And now all you're doing is basically saying
that the features produced by the teacher network
and the student network should be very similar.
That's it.
There is no notion of a negative anymore.
And that's it.
So it's all about similarity maximization
between these two features.
And so all I need to now do is figure out
how to have these two sorts of parallel networks,
a student network and a teacher network.
And basically, researchers have figured out
very cheap methods to do this.
So you can actually have for free really
two types of neural networks.
They're kind of related, but they're different enough
that you can actually basically have
a learning problem set up.
So you can ensure that they always remain different enough
so that the thing doesn't collapse
into something boring.
Exactly.
So the main sort of enemy of self-supervised learning,
any kind of similarity maximization technique is collapse.
It's a collapse means that you learn
the same feature representation
for all the images in the world,
which is completely useless.
Everything is a banana.
Everything is a banana.
Everything is a cat.
Everything is a car.
Yeah.
And so all we need to do is basically come up
with ways to prevent collapse.
Contrasted learning is one way of doing it.
And then, for example, clustering or self-distillation
or other ways of doing it.
We also had a recent paper where we used
decorrelation between two sets of features
to prevent collapse.
So that's inspired a little bit by Horace Barlow's
neuroscience principles.
By the way, I should comment that whoever counts
the number of times the word banana, apple, cat, and dog
were using this conversation wins the internet.
I wish you luck.
What is SWAV and the main improvement proposed
in the paper on supervised learning of visual features
by contrasting cluster assignments?
SWAV basically is a clustering-based technique,
which is, again, the same thing, for self-supervised learning
in vision, where we have two crops.
And the idea, basically, is that you
want the features from these two crops of an image
to lie in the same cluster.
And basically, crops that are coming from different images
to be in different clusters.
Now, typically, if you were to do this clustering,
you would perform clustering offline.
What that means is, if you have a data set of n examples,
you would run over all of these n examples,
get features for them, perform clustering,
so basically get some clusters, and then
repeat the process again.
So this is offline, basically, because I
need to do one pass through the data to compute its clusters.
SWAV is basically just a simple way of doing this online.
So as you're going through the data,
you're actually computing these clusters online.
And so, of course, there is a lot of tricks
involved in how to do this in a robust manner
without collapsing.
But this is this key idea to it.
Is there a nice way to say what is the key methodology
of the clustering that enables that?
Right, so the idea basically is that when you have n samples,
we assume that we have access to, like, there are always
k clusters in a data set.
k is a fixed number.
So for example, k is 3,000.
And so when you look at any sort of small number of examples,
all of them must belong to one of these k clusters.
And we impose this equi partition constraint.
What this means is that basically your entire set of n samples
should be equally partitioned into k clusters.
So all your k clusters have equal contribution
to these n samples.
And this ensures that we never collapse.
So collapse can be viewed as a way in which all samples belong
to one cluster.
So if all features become the same,
then you have basically just one mega cluster.
You don't even have, like, 10 clusters or 3,000 clusters.
So SWAP basically ensures that at each point,
all these 3,000 clusters are being
used in the clustering process.
And that's it.
Basically, just figure out how to do this online.
And again, basically, just make sure
that two crops from the same image
belong to the same cluster, and others don't.
And the fact they have a fixed k makes things simpler.
Fixed k makes things simpler.
Our clustering is not, like, really hard clustering.
It's soft clustering.
So basically, you can be 0.2 to cluster number 1
and 0.8 to cluster number 2.
So it's not really hard.
So essentially, even though we have, like, 3,000 clusters,
we can actually represent a lot of clusters.
What is SEAR, S-E-E-R?
And what are the key results and insights in the paper,
self-supervised, pre-training of visual features in the wild?
What is this big, beautiful SEAR system?
SEAR, so I'll first go to SWAP, because SWAP is actually, like,
one of the key components for SEAR.
So SWAP was, when we use SWAP, it was demonstrated on ImageNet.
So typically, like, self-supervised methods,
the way we sort of operate is, like, in the research community,
we kind of cheat.
So we take ImageNet, which, of course,
I talked about as having lots of labels.
And then we throw away the labels,
like, throw away all the hard work that went behind, basically,
the labeling process.
And we pretend that it is self-unsupervised.
But the problem here is that we have,
like, when we collected these images, the ImageNet dataset
has a particular distribution of concepts, right?
So these images are very curated.
And what that means is these images, of course,
belong to a certain set of noun concepts.
And also, ImageNet has this bias that all images contain
an object which is, like, very big,
and it's typically in the center.
So when you're talking about a dog, it's a well-framed dog.
It's towards the center of the image.
So a lot of the data augmentation,
a lot of the sort of hidden assumptions in self-supervised
learning actually really exploit this bias of ImageNet.
And so, I mean, a lot of my work, a lot of work
from other people always uses ImageNet sort of as the benchmark
to show the success of self-supervised learning.
So you're implying that there's particular limitations
to this kind of dataset?
Yes, I mean, it's basically because our data augmentations
that we designed, like, all the data augmentations
that we designed for self-supervised learning
and vision are kind of overfitted to ImageNet.
But you're saying a little bit hard-coded in, like, the cropping?
Exactly.
The cropping parameters, the kind of lighting that we're using,
the kind of blurring that we're using.
Yeah, but you would, for a more in-the-wild dataset,
you would need to be cleverer, more careful in setting
the range of parameters and those kinds of things.
So for SEER, our main goal was to fold one, basically,
to move away from ImageNet for training.
So the images that we used were uncurated images.
Now, there's a lot of debate whether they're actually
curated or not, but I'll talk about that later.
But the idea was, basically, these
are going to be random internet images
that we're not going to filter out
based on, like, particular categories.
So we did not say that, oh, images that belong to dogs
and cats should be the only images that come in this dataset.
Banana.
And basically, other images should be thrown out.
So we didn't do any of that.
So these are random internet images.
And, of course, it also goes back to, like, the problem
of scale that you talked about.
So these were basically about a billion or so images.
And for Context ImageNet, the ImageNet version that we used
was 1 million images earlier.
So this is basically going, like, three orders of magnitude
more.
The idea was, basically, to see if we
can train a very large convolutional model
in a self-supervised way on this uncurated but really large
set of images.
And how well would this model do?
So is self-supervised learning really
the overfit to ImageNet, or can it actually work in the wild?
And it was also out of curiosity,
what kind of things will this model learn?
Will it actually be able to still figure out, you know,
different types of objects and so on?
Would there be particular kinds of tasks
it would actually do better than an ImageNet trained model?
And so for CIR, one of our main findings
was that we can actually train very large models
in a completely self-supervised way
on lots of internet images without really necessarily
filtering them out, which was in itself a good thing.
Because it's a fairly simple process, right?
So you get images which are uploaded,
and you basically can immediately use them to train
a model in an unsupervised way.
You don't really need to sit and filter them out.
These images can be cartoons, these can be memes,
these can be actual pictures uploaded by people.
And you don't really care about what these images are.
You don't even care about what concepts they contain.
So this was a very sort of simple setup.
What image selection mechanism would you say
is there inherent in some aspect of the process?
So you're kind of implying that there's almost none.
But what is there, would you say, if you want to introspect?
Right, so it's not like uncurated can basically,
like one way of imagining uncurated
is basically you have cameras that can take pictures
at random viewpoints.
When people upload pictures to the internet,
they are typically going to care about the framing of it.
They're not going to upload, say,
the picture of a zoomed-in wall, for example.
Well, when we say internet, do we mean social networks?
Yes.
OK.
So these are not going to be pictures of a zoomed-in table
or a zoomed-in wall.
So it's not really completely uncurated
because people do have their photographer's bias,
where they do want to keep things towards the center
a little bit, or really have nice-looking things
and so on in the picture.
So that's the kind of bias that typically exists
in this data set.
And also the user-based, right?
They're not going to get lots of pictures
from different parts of the world,
because there are certain parts of the world where people may not
actually be uploading a lot of pictures to the internet,
or may not even have access to a lot of internet.
So this is a giant data set and a giant neural network.
I don't think we've talked about what architectures work well
for SSL, for self-supervised learning.
For Seer and for Swab, we were using convolutional networks.
But recently in a work called Dino,
we've basically started using transformers for vision.
Both seem to work really well, con nets and transformers.
And depending on what you want to do,
you might choose to use a particular formulation.
So for Seer, it was a con net.
It was particularly a reg net model,
which was also a work from Facebook.
Reg nets are really good when it comes to compute
versus accuracy.
So because it was a very efficient model,
compute and memory-wise efficient,
and basically it worked really well in terms of scaling.
So we used a very large reg net model
and trained it on a billion images.
Can you maybe quickly comment on what reg nets are?
It comes from this paper, Designing Network Design Spaces.
It's just a super interesting concept
that emphasizes how to create efficient neural networks,
large neural networks.
So one of the key takeaways from this paper,
which the authors, whenever you hear them present this work,
they keep saying is a lot of neural networks
are characterized in terms of flops, right?
Flops basically being the floating point operations.
And people really love to use flops to say,
this model is really computationally heavy,
or our model is computationally cheap, and so on.
Now it turns out that flops are really not a good indicator
of how well a particular network is,
like how efficient it is really.
And what a better indicator is, is the activation,
or the memory that is being used by this particular model.
And so Designing, one of the key findings from this paper
was basically that you need to design network families,
or neural network architectures,
that are actually very efficient in the memory space as well,
not just in terms of pure flops.
So reg net is basically a network architecture family
that came out of this paper,
that is particularly good at both flops
and the sort of memory required for it.
And of course, it builds upon like earlier work,
like ResNet being like the sort of more popular inspiration
for it, where you have residual connections.
But one of the things in this work is basically
they also use like squeeze excitation blocks.
So it's a lot of nice sort of technical innovation
in all of this from prior work,
and a lot of the ingenuity of these particular authors
in how to combine these multiple building blocks.
But the key constraint was optimized
for both flops and memory when you're basically doing this.
Don't just look at flops.
And that allows you to what have sort of have very large
networks through this process can optimize for low,
like for efficiency, for low memory.
Also in just in terms of pure hardware,
they fit very well on GPU memory.
So they can be like really powerful neural network
architectures with lots of parameters, lots of flops,
but also because they're like efficient in terms of
the amount of memory that they're using,
you can actually fit a lot of these on,
like you can fit a very large model
on a single GPU, for example.
Would you say that the choice of architecture matters more
than the choice of maybe data augmentation techniques?
Is there a possibility to say what matters more?
You kind of imply that you can probably go really far
with just using basic convenants.
All right, I think data and data augmentation,
the algorithm being used for the cell supervised training
matters a lot more than the particular kind of architecture.
With different types of architecture,
you'll get different properties in the resulting
sort of representation, but really,
I mean, the secret sauce is in the data augmentation
and the algorithm being used to train them.
The architectures, I mean, at this point,
a lot of them perform very similarly,
depending on the particular task that you care about,
they have certain advantages and disadvantages.
Is there something interesting to be said
about what it takes with Sears
to train a giant neural network?
You're talking about a huge amount of data,
a huge neural network.
Is there something interesting to be said
of how to effectively train something like that fast?
Lots of GPUs.
OK, so this.
I mean, so the model was like a billion parameters,
and it was trained on the billion images.
So basically the same number of parameters
as the number of images, and it took a while.
I don't remember the exact number, it's in the paper,
but it took a while.
I guess I'm trying to get at is when
you're thinking of scaling this kind of thing.
I mean, one of the exciting possibilities
of self-supervised learning is the several orders
of magnitude scaling of everything,
both the neural network and the size of the data.
And so the question is, do you think
there are some interesting tricks
to do large-scale distributed compute,
or is that really outside of even deep learning?
That's more about hardware engineering.
I think more and more there is this lot of systems
are designed basically taking into account
the machine learning needs.
So because whenever you're doing this kind of distributed
training, there is a lot of intercommunication
between nodes, so gradients or the model parameters
are being passed.
So you really want to minimize communication costs
when you really want to scale these models up.
You want basically to be able to do as much,
as limited amount of communication as possible.
So currently a dominant paradigm
is synchronized training.
So essentially, after every gradient step,
all you basically have a synchronization
step between all the compute chips that you're going on with.
I think asynchronous training was popular,
but it doesn't seem to perform as well.
But in general, I think that's sort of the,
I guess it's outside my scope as well.
But the main thing is minimize the amount of synchronization
steps that you have.
That has been the key takeaway, at least in my experience.
The others, I have no idea about how to design the chip.
Yeah, there's very few things that I see Jim Keller's eyes
light up as much as talking about giant computers doing
like that fast communication that you're talking to well
when they're training machine learning systems.
What is Vistle, the ISSL, the PyTorch-based SSL library?
What are the use cases that you might have?
Vistle basically was born out of a lot of us at Facebook
doing the self-supervised learning research.
So it's a common framework in which we
have a lot of self-supervised learning methods
implemented for vision.
It has in itself a benchmark of tasks
that you can evaluate the self-supervised representation
zone.
So the use case for it is basically
for anyone who's either trying to evaluate their self-supervised
model or train their self-supervised model
or a researcher who's trying to build a new self-supervised
technique.
So it's basically supposed to be all of these things.
So as a researcher, before Vistle, for example,
or when we started doing this work fairly seriously
at Facebook, it was very hard for us
to go and implement every self-supervised learning
model, test it out in a consistent manner.
The experimental setup was very different
across different groups.
Even when someone said that they were reporting image net
accuracy, it could mean lots of different things.
So with Vistle, we tried to really standardize that
as much as possible.
And it was a paper we did in 2019 just about benchmarking.
And so Vistle basically builds upon this kind of work
that we did about benchmarking.
And then every time we try to come up with a self-supervised
learning method, a lot of us try to push that into Vistle
as well, just so that it basically
is the central piece where a lot of these methods can decide.
Just out of curiosity, people may be certainly outside
of Facebook, but just researchers,
or just even people that know how to program in Python
and know how to use PyTorch, what would be the use case?
What would be a fun thing to play around with Vistle on?
What's a fun thing to play around with self-supervised
learning on, would you say?
Is there a good Hello World program?
Is it always about big size that's important to have?
Or is there fun, little, smaller case playgrounds
to play around with?
So we're trying to push something towards that.
I think there are a few setups out there,
but nothing super standard on the smaller scale.
I mean, ImageNet in itself is actually pretty big also.
So that is not something which is feasible for a lot of people.
But we are trying to push up with smaller use cases.
The thing is, at a smaller scale,
a lot of the observations or a lot of the algorithms that
work don't necessarily translate into the medium
or the larger scale.
So it's really tricky to come up with a good small scale
setup where a lot of your empirical observations will
really translate to the other setup.
So it's been really challenging.
I've been trying to do that for a little bit as well,
because it does take time to train stuff on ImageNet.
It does take time to train on more images.
But pretty much every time I've tried to do that,
it's been unsuccessful, because all the observations
I draw from my set of experiments on a smaller data set
don't translate into ImageNet or don't translate
into another sort of data set.
So it's been hard for us to figure this one out,
but it's an important problem.
So there's this really interesting idea
of learning across multiple modalities.
You have a CVPR 2021 best paper candidate titled
Audiovisual Instance Discrimination
with Cross-Modal Agreement.
What are the key results, insights in this paper,
and what can you say in general about the promise and power
of multimodal learning?
For this paper, it actually came as a little bit of a shock
to me at how well it worked, so I can describe what
the problem setup was.
So it's been used in the past by lots of folks,
like for example, Andrew Owens from MIT,
Aliyo Shai from Berkeley, Andrew Zisserman from Oxford.
So a lot of these people have been showing results in this.
Of course, I was aware of this result,
but I wasn't really sure how well it would work in practice
for other downstream tasks.
So the results kept getting better,
and I wasn't sure if a lot of our insights from self-supervised
learning would translate into this multimodal learning
problem.
So multimodal learning is when you have multiple modalities.
And that's not even possible.
OK, so the particular modalities
that we worked on in this work were audio and video.
So the idea was basically if you have a video,
you have its corresponding audio track,
and you want to use both of these signals, the audio signal
and the video signal to learn a good representation for video
and good representation for audio.
Like this podcast.
Like this podcast, exactly.
So what we did in this work was basically
train two different neural networks, one on the video
signal, one on the audio signal.
And what we wanted is basically the features
that we get from both of these neural networks
should be similar.
So it should basically be able to produce the same kinds
of features from the video and the same kinds of features
from the audio.
Now, why is this useful?
Well, for a lot of these objects that we have,
there is a characteristic sound.
So trains, when they go by, they make a particular kind of sound.
Boats make a particular kind of sound.
People, when they're jumping around, will shout, whatever.
Bananas don't make a sound.
So well, you can't learn anything about bananas there.
Or when humans mention bananas.
Well, yes.
When they say the word banana, then you can't
trust basically anything that comes out of a human's mouth
as a source, that source of audio is useless.
So the typical use case is basically like, for example,
someone playing a musical instrument.
So guitars have a particular kind of sound and so on.
So because a lot of these things are correlated,
the idea in multimodal learning is
to take these two kinds of modalities, video and audio,
and learn a common embedding space, a common feature
space where both of these related modalities
can basically be close together.
And again, you use contrastive learning for this.
So in contrastive learning, basically the video
and the corresponding audio are positives.
And you can take any other video or any other audio,
and that becomes a negative.
And so basically that's it.
It's just a simple application of contrastive learning.
The main sort of finding from this work for us
was basically that you can actually
learn very, very powerful feature representations, very,
very powerful video representations.
So you can learn the sort of video network
that we ended up learning can actually
be used for downstream, for example, recognizing human actions,
or recognizing different types of sounds, for example.
So this was sort of the key finding.
Can you give kind of an example of a human action,
or just so we can build up intuition
of what kind of thing?
Right, so there is this data set called kinetics,
for example, which has like 400 different types
of human actions.
So people jumping, people doing different kinds of sports,
or different types of swimming, so different strokes
and swimming, golf, and so on.
So there are just different types of actions right there.
And the point is this kind of video network
that you learn in a self-supervised way
can be used very easily to kind of recognize
these different types of actions.
It can also be used for recognizing
different types of objects.
And what we did is we tried to visualize
whether the network can figure out where
the sound is coming from.
So basically give it a video, and basically play of, say,
of a person just strumming a guitar,
but of course, there is no audio in this.
And now you give it the sound of a guitar.
And you ask, like, basically try to visualize
where the network is, where the network thinks
the sound is coming from.
And it can kind of basically draw, like, when you visualize it,
you can see that it's basically focusing on the guitar.
Yeah, that's surreal.
And the same thing, for example,
for certain people's voices, like famous celebrities' voices,
it can actually figure out where their mouth is.
So it can actually distinguish different people's voices,
for example, a little bit as well.
Without that ever being annotated in any way.
Right.
So this is all what it had discovered.
We never pointed out that this is a guitar,
and this is the kind of sound it produces.
It can actually naturally figure that out,
because it's seen so many correlations of this sound
coming with this kind of, like, an object,
that it basically learns to associate this sound
with this kind of an object.
Yeah, that's really fascinating, right?
That's really interesting.
And so the idea with this kind of network
is then you then fine-tune it for a particular task.
Right.
So this is forming, like, a really good knowledge base
within a neural network, based on which you could then
train a little bit more to accomplish a specific task well.
Exactly.
So you don't need a lot of videos of humans
doing actions annotated.
You can just use a few of them to basically get your recognition.
How much insight do you draw from the fact
that you can figure out where the sound is coming from?
I'm trying to see.
So that's kind of very, it's very CVPR-beautiful, right?
It's a cool little insight.
I wonder how profound that is.
Does it speak to the idea that multiple modalities are somehow
much bigger than the sum of their parts?
Or is it really, really useful to have multiple modalities?
Or is it just that cool thing that there's parts of our world
that can be revealed effectively through multiple modalities?
But most of it is really all about vision
or about one of the modalities.
I would say a little tending more towards the second part.
So most of it can be figured out with one modality,
but having an extra modality always helps you.
So in this case, for example, one thing is when you're,
if you observe someone cutting something
and you don't have any sort of sound there,
whether it's an apple or whether it's an onion,
it's very hard to figure that out.
But if you hear someone cutting it,
it's very easy to figure it out.
Because apples and onions make a very different kind
of characteristic sound when they're cut.
So you really figure this out based on audio.
It's much easier.
So your life will become much easier
when you have access to different kinds of modalities.
And the other thing is, so I like to relate it in this way.
It may be completely wrong.
But the distributional hypothesis in NLP,
where context basically gives kind of meaning to that word.
Sound kind of does that too.
So if you have the same sound, so that's the same context,
across different videos, you're very likely to be observing
the same kind of concept.
So that's the kind of reason why it figures out
the guitar thing, right?
It observed the same sound across multiple different videos.
And it figures out maybe this is the common factor
that's actually doing it.
I wonder, I used to have this argument with my dad a bunch
for creating general intelligence,
whether smell is an important,
like if that's important sensory information.
Mostly we're talking about like falling in love
with an AI system.
And for him, smell and touch are important.
And I was arguing that it's not at all,
it's nice and everything,
but like you can fall in love with just language really,
but voice is very powerful and vision is next
and smell is not that important.
Can I ask you about this process of active learning?
You mentioned interactivity.
Is there some value within the self-supervised
learning context to select parts of the data
in intelligent ways such that they would most benefit
the learning process?
So I think so.
I mean, I know I'm talking to an active learning fan here,
so of course I know the answer.
First you were talking bananas
and now you're talking about active learning, I love it.
I think Yana Koon told me that active learning
is not that interesting.
And I think back then I didn't want to argue with him too much,
but when we talk again,
we're gonna spend three hours arguing about active learning.
My sense was you can go extremely far with active learning,
perhaps farther than anything else.
Like to me, there's this kind of intuition
that similar to data augmentation,
you can get a lot from the data,
from intelligent optimized usage of the data.
I'm trying to speak generally in such a way
that includes data augmentation and active learning,
that there's something about maybe interactive exploration
of the data that at least this part
of the solution to intelligence, like an important part.
I don't know what your thoughts are
on active learning in general.
I actually really like active learning.
So back in the day we did this largely ignored
CVPR paper called Learning by Asking Questions.
So the idea was basically you would train an agent
that would ask a question about the image,
it would get an answer,
and basically then it would update itself,
it would see the next image,
it would decide what's the next hardest question
that I can ask to learn the most.
And the idea was basically because it was being smart
about the kinds of questions it was asking,
it would learn in fewer samples,
it would be more efficient at using data.
And we did find to some extent
that it was actually better than randomly asking questions.
Kind of weird thing about active learning is
it's also a chicken and egg problem
because when you look at an image,
to ask a good question about the image,
you need to understand something about the image.
You can't ask a completely arbitrarily random question,
it may not even apply to that particular image.
So there is some amount of understanding or knowledge
that basically keeps getting built
when you're doing active learning.
So I think active learning in by itself is really good.
And the main thing we need to figure out is basically
how do we come up with a technique
to first model what the model knows
and also model what the model does not know.
I think that's the sort of beauty of it, right?
Because when you know that there are certain things
that you don't know anything about,
asking a question about those concepts
is actually going to bring you the most value.
And I think that's the sort of key challenge.
Now self-supervised learning by itself,
like selecting data for it and so on,
that's actually really useful.
But I think that's a very narrow view
of looking at active learning, right?
If you look at it more broadly,
it is basically about if the model has a knowledge
about end concepts and it is weak basically
about certain things.
So it needs to ask questions either to discover new concepts
or to basically like increase its knowledge
about these end concepts.
So at that level, it's a very powerful technique.
I actually do think it's going to be really useful.
Even in like simple things such as like data labeling,
it's super useful.
So here is like one simple way
that you can use active learning.
For example, you have your self-supervised model,
which is very good at predicting similarities
and dissimilarities between things.
And so if you label a picture as basically say a banana,
now you know that all the images
that are very similar to this image
are also likely to contain bananas.
So probably when you want to understand
what else is a banana,
you're not going to use these other images.
You're actually going to use an image
that is not completely dissimilar,
but somewhere in between,
which is not super similar to this image,
but not super dissimilar either.
And that's going to tell you a lot more
about what this concept of a banana is.
So that's kind of a heuristic.
I wonder if it's possible to also learn ways
to discover the most likely,
the most beneficial image.
So not just looking a thing that's somewhat similar
to a banana, but not exactly similar,
but have some kind of more complicated learning system,
like learned discovery mechanism
that tells you what image to look for.
Like how, yeah, like actually in a self-supervised way,
learning strictly a function that says,
is this image going to be very useful to me,
given what I currently know?
I think there is a lot of synergy there.
It's just, I think, yeah, it's going to be explored.
I think very much related to that.
I kind of think of what Tesla autopilot is doing
at currently as kind of active learning.
There's something that Andre Capati and their team
are calling data engine.
So you're basically deploying a bunch of instantiations
of a neural network into the wild,
and they're collecting a bunch of edge cases
that are then sent back for annotation, for particular,
and edge cases as defined as near failure
or some weirdness on a particular task
that's then sent back.
It's that not exactly a banana,
but almost a banana cases sent back for annotation.
And then there's this loop that keeps going
and you keep retraining and retraining.
And the active learning step there,
or whatever you want to call it,
is the cars themselves that are sending you back the data,
like what the hell happened here?
This was weird.
What are your thoughts about that sort of deployment
of neural networks in the wild?
Another way to ask a question for first is your thoughts,
and maybe if you want to comment,
is there applications for autonomous driving,
like computer vision-based autonomous driving,
applications of self-supervised learning
in the context of computer vision-based autonomous driving?
So I think so.
I think for self-supervised learning
to be used in autonomous driving,
there are lots of opportunities.
And just like pure consistency in predictions
is one way, right?
So because you have this nice sequence of data
that is coming in, a video stream of it,
associated of course with the actions
that say the car took,
you can form a very nice predictive model
of what's happening.
So for example, like all the way,
like one way possibly in which how they're figuring out
what data to get labeled is basically
through prediction uncertainty, right?
So you predict that the car was going to turn right.
So this was the action that was going to happen,
say in the shadow mode,
and now the driver turned left.
And this is a really big surprise.
So basically by forming these good predictive models,
you are, I mean, these are kind of self-supervised models,
right?
Prediction models are basically being trained
just by looking at what's going to happen next
and asking them to predict what's going to happen next.
So I would say this is really
like one use of self-supervised learning.
It's a predictive model
and you're learning a predictive model
basically just by looking at what data you have.
Is there something about that active learning context
that you find insights from?
Like that kind of deployment of the system,
seeing cases where it doesn't perform as you expected
and then retraining the system based on that?
I think that, I mean, that really resonates with me.
It's super smart to do it that way.
Because I mean, the thing is with any kind of like
practical system like autonomous driving,
there are those edge cases
that are the things that are actually the problem, right?
I mean, highway driving or like freeway driving
has basically been like,
there has been a lot of success in that particular part
of autonomous driving for a long time.
I would say like since the 80s or something.
Now, the point is all these failure cases
are the sort of reason why autonomous driving
hasn't become like super, super mainstream
available like in every possible car right now.
And so basically by really scaling this problem out
by really trying to get all of these edge cases out
as quickly as possible.
And then just like using those to improve your model,
that's super smart.
And prediction uncertainty to do that
is like one really nice way of doing it.
Let me put you on the spot.
So we mentioned offline Jitendra.
He thinks that the Tesla computer vision approach
or really any approach for autonomous driving
is very far away.
How many years away,
if you have to bet all your money on it,
are we dissolving autonomous driving
with this kind of computer vision only
machine learning based approach?
Okay, so what does solving autonomous driving mean?
Does it mean solving it in the US?
Does it mean solving it in India?
Because I can tell you that very different types
of driving happening.
Not India, not Russia.
In the United States autonomous,
so what solving means is when the car says it has control,
it is fully liable.
You can go to sleep, it's driving by itself.
So this is highway and city driving,
but not everywhere, but mostly everywhere.
And it's, let's say significantly better,
like say five times less accidents than humans.
It's sufficiently safer such that the public feels
like that transition is enticing beneficial
both for our safety and financial
and all those kinds of things.
Okay, so first disclaimer,
I'm not an expert in autonomous driving.
So let me put it out there.
I would say like at least five to 10 years.
This is, this would be my, like guess from now.
The, yeah, I'm actually very impressed.
Like when I sat in a friend's Tesla recently
and of course like looking, so it can,
on the screen it basically shows all the detections
and everything the car is doing as you're driving by.
And that's super distracting for me as a person
because all I keep looking at is like the bounding boxes
and the cars, it's tracking and it's really impressive.
Like especially when it's raining
and it's able to do that,
that was the most impressive part for me.
It's actually able to get through rain and do that.
And one of the reasons why like a lot of us believed
and I would put myself in that category
is LiDAR based sort of technology
for autonomous driving was the key driver, right?
So Waymo was using it for the longest time.
And Tesla then decided to go this completely other route
that oh, we're not going to even use LiDAR.
So their initial system I think was camera and radar based
and now they're actually moving to a completely
like vision based system.
And so that was just like, it sounded completely crazy.
LiDAR is very useful in cases where you have low visibility.
Of course it comes with its own set of complications.
But now to see that happen in like on a live Tesla,
that basically just proves everyone wrong,
I would say in a way.
And that's just working really well.
I think there were also like a lot of advancements
in camera technology.
Now there were like, I know at CMU when I was there,
there was a particular kind of camera
that had been developed that was really good
at basically low visibility settings.
So like lots of snow and lots of rain,
it could actually still have a very reasonable visibility.
And I think there are lots of these kinds of innovations
that will happen on the sensor side itself,
which is actually going to make this very easy in the future.
And so maybe that's actually why I'm more optimistic
about vision based self like autonomous driving.
It's gonna call it self supervised driving,
but vision based autonomous driving,
that's the reason I'm quite optimistic about it.
Because I think there are going to be lots of these advances
on the sensor side itself.
So acquiring this data,
we're actually going to get much better about it.
And then of course, when once we're able to scale out
and get all of these edge cases in,
as like Andre described,
I think that's going to make us go very far away.
Yeah, so it's funny, I'm very much with you
on the five to 10 years, maybe 10 years,
but you made it, I'm not sure how you made it sound,
but for some people that might seem like really far away.
And then for other people, it might seem like very close.
There's a lot of fundamental questions
about how much game theory is in this whole thing.
So like how much is this simply collision avoidance problem?
And how much of it is,
you're still interacting with other humans in the scene,
and you're trying to create an experience that's compelling.
So you want to get from point A to point B quickly,
you want to navigate the scene in a safe way,
but you also want to show some level of aggression,
because well, certainly this is why you're screwed in India
because you have to show aggression.
Or Jersey, or New Jersey.
Or Jersey, right?
So like, or New York, or basically any major city,
but I think it's probably Elon that I talked the most
about this, which is a surprise to the level
of which they're not considering human beings
as a huge problem in this, as a source of problem.
Like the driving is fundamentally a robot on robot
versus the environment problem,
versus like a, you know,
you can just consider humans not part of the problem.
I used to think humans are almost certainly
have to be modeled really well.
Pedestrians and cyclists and humans inside of the cars,
you have to have like mental models for them.
You cannot just see it as objects.
But more and more, it's like the,
it's the same kind of intuition-breaking thing
that self-supervised learning does,
which is, well, maybe through the learning,
you'll get all the human,
like human information you need, right?
Like maybe you'll get it just with enough data.
You don't need to have explicit good models
of human behavior.
Maybe you get it through the data.
So, I mean, my skepticism also just knowing
a lot of automotive companies
and how difficult it is to be innovative,
I was skeptical that they would be able at scale
to convert the driving scene across the world
into digital form such that
you can create this data engine at scale.
And the fact that Tesla is at least getting there
or are already there makes me think that
it's now starting to be coupled
to the self-supervised learning vision,
which is like, if that's gonna work,
if through purely this process you can get really far,
then maybe you can solve driving that way.
I don't know.
I tend to believe we don't give enough credit
to the how amazing humans are both at driving
and at supervising autonomous systems.
And also we don't, this is, I wish we were,
I wish there was much more driver sensing inside Teslas
and much deeper consideration of human factors,
like understanding psychology and drowsiness
and all those kinds of things.
When the car does more and more of the work,
how to keep utilizing the little human supervision
that are needed to keep this whole thing safe.
I mean, it's a fascinating dance of human robot interaction.
To me, autonomous driving for a long time
is a human robot interaction problem.
It is not a robotics problem or computer vision problem.
Like you have to have a human in the loop.
Which is why I think it's 10 years plus,
but I do think there'll be a bunch of cities and contexts
where you're restricted, it will work really, really damn well.
So I think for me, it's five if I'm being optimistic
and it's going to be five for a lot of cases
and 10 plus, yeah, I agree with you.
10 plus, basically, if we want to recover
most of the contiguous United States or something.
Oh, interesting.
So my optimistic is five and pessimistic is 30.
30.
I have a long tail on this one.
I haven't watched enough driving videos.
I've watched enough pedestrians to think like,
we may be, like there's a small part of me still,
not a small, like a pretty big part of me
that thinks we will have to build AGI to solve driving.
Oh well.
Like there's something to me like,
because humans are part of the picture,
deeply part of the picture.
And also human society is part of the picture
in that human life is at stake.
Anytime a robot kills a human,
it's not clear to me that that's not a problem
that machine learning will also have to solve.
Like it has to, you have to integrate that
into the whole thing, just like Facebook
or social networks, one thing is to say
how to make a really good recommender system.
And then the other thing is to integrate
into that recommender system,
all the journalists that will write articles
about that recommender system.
Like you have to consider the society
within which the AI system operates.
And in order to, and like politicians too,
this is there's regulatory stuff for autonomous driving.
It's kind of fascinating that the more successful
your AI system becomes, the more it gets integrated
in society and the more precious politicians
and the public and the clickbait journalists
and all the different fascinating forces
of our society start acting on it.
And then it's no longer how good you are
at doing the initial task,
it's also how good you are at navigating human nature,
which is a fascinating space.
What do you think are the limits of deep learning?
If you allow me, we'll zoom out a little bit
into the big question of artificial intelligence.
You said dark matter of intelligence
is self-supervised learning, but there could be more.
What do you think the limits of self-supervised learning
and just learning in general, deep learning are?
I think like for deep learning in particular,
because self-supervised learning is I would say
a little bit more vague right now.
So I wouldn't, like for something that's so vague,
it's hard to predict what its limits are going to be.
But like I said, I think anywhere you want to interact
with humans, self-supervised learning kind of hits a boundary
very quickly because you need to have an interface
to be able to communicate with the human.
So really like if you have just like vacuous concepts
or like just like nebulous concepts discovered
by a network, it's very hard to communicate those
with the human without like inserting some kind
of human knowledge or some kind of like human bias there.
In general, I think for deep learning,
the biggest challenge is just like data efficiency.
Even with self-supervised learning,
even with anything else, if you just see
a single concept once, like one image of a,
like I don't know, whatever you want to call it,
like any concept, it's really hard for these methods
to generalize by looking at just one or two samples of things.
And that has been a real challenge.
And I think that's actually why like these edge cases,
for example, for Tesla are actually that important.
Because if you see just one instance of the car failing,
and if you just annotate that and you get that
into your data set, you have like very limited guarantee
that it's not going to happen again.
And you're actually going to be able to recognize
this kind of instance in a very different scenario.
So like when it was snowing, so you got that thing labeled
when it was snowing, but now when it's raining,
you're actually not able to get it.
Or you basically have the same scenario
in a different part of the world.
So the lighting was different or so on.
So it's just really hard for these models,
like deep learning, especially to do that.
What's your intuition?
How do we solve Henry and Digit Recognition problem
when we only have one example for each number?
It feels like humans are using something like learning.
Right.
I think it's, we are good at transferring knowledge
a little bit.
We are just better at like, for a lot of these problems
where we are generalizing from a single sample
or recognizing from a single sample.
We are using a lot of our own domain knowledge
and a lot of our like inductive bias
into that one sample to generalize it.
So I've never seen you write the number nine, for example.
And if you were to write it, I would still get it.
And if you were to write a different kind of alphabet
and like write it in two different ways,
I would still probably be able to figure out
that these are the same two characters.
It's just that I have been very used to seeing
Henry and Digits in my life.
The other sort of problem with any deep learning system
or any kind of machine learning system is like,
it's guarantees, right?
There are no guarantees for it.
Now you can argue that humans also don't have any guarantees.
Like there is no guarantee that I can recognize a cat
in every scenario.
I'm sure there are going to be lots of cats
that I don't recognize, lots of scenarios
in which I don't recognize cats in general.
But I think from like, from just a sort of application
perspective, you do need guarantees, right?
We call these things algorithms.
Now algorithms, like traditional CS algorithms
have guarantees, sorting is a guarantee.
If you were to call sort on a particular array of numbers,
you are guaranteed that it's going to be sorted,
otherwise it's a bug.
Now for machine learning, it's very hard to characterize this.
We know for a fact that like a cat recognition model
is not going to recognize cats, every cat in the world
in every circumstance.
I think most people would agree with that statement.
But we are still okay with it.
We still don't call this as a bug.
Whereas in traditional computer science
or traditional science, like if you have this kind
of failure case existing, then you think of it
as like something is wrong.
I think there is this sort of notion
of nebulous correctness for machine learning.
And that's something we just need to be very comfortable with.
And for deep learning or like for a lot
of these machine learning algorithms,
it's not clear how do we characterize
this notion of correctness.
I think limitation in our understanding
or at least a limitation in our phrasing of this.
And if we were to come up with better ways
to understand this limitation,
then it would actually help us a lot.
Do you think there's a distinction between
the concept of learning and the concept of reasoning?
Do you think it's possible for neural networks to reason?
So I think of it slightly differently.
So for me, learning is whenever I can
like make a snap judgment.
So if you show me a picture of a dog,
I can immediately say it's a dog.
But if you give me like a puzzle, you know,
like whatever a Goldberg machine of like
things going to happen, then I have to reason.
Because I've never, it's a very complicated setup.
I've never seen that particular setup
and I really need to draw and like imagine in my head
what's going to happen to figure it out.
So I think, yes, neural networks are really good
at recognition, but they're not very good at reasoning.
Because they're like, if they have seen something before
or seen something similar before,
they're very good at making those sort of snap judgments.
But if you were to give them a very complicated thing
that they've not seen before,
they have very limited ability right now
to compose different things.
Like, oh, I've seen this particular part before.
I've seen this particular part before.
And now probably like this is how they're going
to work in tandem.
It's very hard for them to come up
with these kinds of things.
Well, there's a certain aspect to reasoning
that you can maybe convert into the process of programming.
And so there's the whole field of program synthesis
and people have been applying machine learning
to the problem of program synthesis.
And the question is, you know, can they,
the step of composition, why can't that be learned?
You know, this step of like building things on top of,
like little intuitions, concepts on top of each other,
can that be learnable?
What's your intuition there?
Or like, I guess similar set of techniques,
do you think they'll be applicable?
So I think it is, of course, it is learnable
because like we are prime examples of machines
that have like, or individuals that have learned this, right?
Like humans have learned this.
So it is, of course, it is a technique
that is very easy to learn.
I think where we are kind of hitting a wall,
basically with like current machine learning
is the fact that when the network learns
all of this information,
we basically are not able to figure out
how well it's going to generalize to an unseen thing.
And we have no, like a priori, no way of characterizing that.
And I think that's basically telling us a lot about,
like a lot about the fact that we really don't know
what this model has learned and how well it's basically,
because we don't know how well it's going to transfer.
There's also a sense in which it feels like
we humans may not be aware of how much like background,
how good our background model is,
how much knowledge we just have slowly building
on top of each other.
It feels like neural networks are constantly
throwing stuff out.
Like you'll do some incredible thing
where you're learning a particular task in computer vision,
you celebrate your state of the art successes
and you throw that out.
Like it feels like it's,
you're never using stuff you've learned
for your future successes in other domains.
And humans are obviously doing that exceptionally well,
still throwing stuff away in their mind,
but keeping certain kernels of truth.
Right, so I think we're like,
continual learning is sort of the paradigm
for listen machine learning.
And I don't think it's a very well explored paradigm.
We have like things in deep learning, for example,
that catastrophic forgetting is like one of the standard things.
The thing basically being that if you teach a network
like to recognize dogs,
and now you teach that same network to recognize cats,
it basically forgets how to recognize dogs.
So it forgets very quickly.
I mean, and whereas a human,
if you were to teach someone to recognize dogs
and then to recognize cats,
they don't forget immediately how to recognize these dogs.
I think that's basically sort of what you're trying to get.
Yeah, I just, I wonder if like the long term memory mechanisms
or the mechanisms that store,
not just memories, but concepts
that allow you to reason and compose concepts.
If those things will look very different than neural networks,
or if you can do that within a single neural network
with some particular sort of architecture quirks,
that seems to be a really open problem.
And of course, I go up and down on that
because there's something so compelling
to the symbolic AI or to the ideas of logic-based
sort of expert systems.
You have like human interpretable facts
that built on top of each other.
It's really annoying, like with self-supervised learning,
that the AI is not very explainable.
Like you can't like understand
all the beautiful thing has learned.
You can't ask it like questions.
But then again, maybe that's a stupid thing
for us humans to want.
Right, I think whenever we try to understand it,
we're putting our own subjective human bias into it.
And I think that's the sort of problem.
With self-supervised learning,
the goal is that it should learn naturally from the data.
So now if you try to understand it,
you are using your own preconceived notions
of what this model has learned.
And that's the problem.
High level question.
What do you think it takes to build a system
with super human, maybe let's say human level
or super human level general intelligence?
We've already kind of started talking about this,
but what's your intuition?
Like does this thing have to have a body?
Does it have to interact richly with the world?
Does it have to have some more human elements
like self-awareness?
I think emotion.
I think emotion is something which is,
like it's not really attributed
typically in standard machine learning.
It's not something we think about.
Like there is NLP, there is vision,
there is no like emotion.
Emotion is never a part of all of this.
And that just seems a little bit weird to me.
I think the reason basically being that
there is surprise and like,
basically emotion is like one of the reasons
emotions arises like what happens
and what you expect to happen, right?
There is like a mismatch between these things.
And so that gives rise like,
I can either be surprised or I can be saddened
or I can be happy and all of this.
And so this basically indicates
that I already have a predictive model in my head
and something that I predicted
or something that I thought was likely to happen.
And then there was something that I observed that happened
that there was a disconnect between these two things.
And that basically is like,
maybe one of the reasons I like,
you have a lot of emotions.
Yeah.
I think, so I talked to people a lot about them
like Lisa Feldman Barrett.
I think that's an interesting concept of emotion,
but I have a sense that emotion primarily
in the way we think about it,
which is the display of emotion
is a communication mechanism between humans.
So it's a part of basically human-to-human interaction.
An important part, but just the part.
So it's like, I would throw it into the full mix
of communication and to me,
communication can be done with objects
that don't look at all like humans.
Okay.
I've seen our ability to anthropomorphize,
our ability to connect with things that look like a Roomba,
our ability to connect.
First of all, let's talk about other biological systems
like dogs, our ability to love things
that are very different than humans.
But they do display emotion, right?
I mean, dogs do display emotion.
So they don't have to be anthropomorphic for them
to like display the kind of emotions that we don't.
Exactly.
So I mean, but then the word emotion starts to lose.
So then we have to be, I guess, specific.
But yeah, so have rich flavorful communication.
Communication, yeah.
Yeah, so like, yes, it's full of emotion.
It's full of wit and humor and moods
and all those kinds of things.
Yeah, so you're talking about like flavor.
Flavor, yeah, okay, let's call it that.
So there's content and then there is flavor
and I'm talking about the flavor.
Do you think it needs to have a body?
Do you think, like to interact with a physical world,
do you think you can understand the physical world
without being able to directly interact with it?
I don't think so, yeah.
I think at some point we will need to bite the bullet
and actually interact with the physical as much as I like
working on like passive computer vision
where I just like sit in my armchair
and look at videos and learn.
I do think that we will need to have some kind of embodiment
or some kind of interaction
to figure out things about the world.
What about consciousness?
You think, how often do you think about consciousness
when you think about your work?
You could think of it as the more simple thing
of self-awareness, of being aware
that you are a perceiving, sensing, acting thing
in this world or you can think about the bigger version
of that which is consciousness,
which is having it feel like something to be that entity,
the subjective experience of being in this world.
So I think of self-awareness a little bit more
than the broader goal of it
because I think self-awareness is pretty critical
for any kind of AGI or whatever you want to call it
that we build because it needs to contextualize
what it is and what role it's playing
with respect to all the other things that exist around it.
I think that requires self-awareness.
It needs to understand that it's an autonomous car, right?
And what does that mean?
What are its limitations?
What are the things that it is supposed to do and so on?
What is its role in some way?
Or I mean, these are the kind of things
that we kind of expect from it, I would say.
And so that's the level of self-awareness
that I would say basically required at least,
if not more than that.
Yeah, I tend to on the emotion side believe
that it has to be able to display consciousness.
Display consciousness, what do you mean by that?
Meaning like for us humans to connect with each other
or to connect with other living entities,
I think we need to feel like in order for us to truly feel
like that there's another being there,
we have to believe that they're conscious.
And so we won't ever connect with something
that doesn't have elements of consciousness.
Now I tend to think that that's easier to achieve
than it may sound.
As we anthropomorphize stuff so hard,
like you have a mug that just like has wheels
and like rotates every once in a while and makes a sound.
I think a couple of days in, especially if you're,
if you don't hang out with humans,
you might start to believe that mug on wheels is conscious.
So I think we anthropomorphize pretty effectively
as human beings.
But I do think that it's in the same bucket
that we'll call emotion.
And show that you're,
I think of consciousness as the capacity to suffer.
And if you're an entity that's able to feel things
in the world and to communicate that to others,
I think that's a really powerful way to interact with humans.
And in order to create an AGI system,
I believe you should be able to richly interact with humans.
Like humans would need to want to interact with you.
Like it can't be like, it's the self-supervised learning
versus like, like the robot shouldn't have to pay you
to interact with me.
So like it should be a natural fun thing.
And then you're going to scale up significantly
how much interaction it gets.
It's the elect surprise,
which they're trying to give me to be a judge on their contest.
Let's see if I want to do that.
But their challenge is to talk to you,
make the human sufficiently interested
that the human keeps talking for 20 minutes.
To Alexa.
To Alexa, yeah.
And right now they're not even close to that
because it just gets so boring when you're like,
when the intelligence is not there,
it gets very not interesting to talk to it.
And so the robot needs to be interesting.
And one of the ways it can be interesting
is display the capacity to love, to suffer.
And I would say that essentially means
the capacity to display consciousness.
Like it is an entity, much like a human being.
Of course, what that really means,
I don't know if that's fundamentally a robotics problem
or some kind of problem that we're not yet even aware.
Like if it is truly a hard problem of consciousness,
I tend to maybe optimistically think it's a,
we can pretty effectively fake it till we make it.
So we can display a lot of human-like elements for a while.
And that will be sufficient to form
really close connections with humans.
What to use the most beautiful idea
in self-supervised learning?
Like when you sit back with, I don't know,
with a glass of wine and armchair
and just at a fireplace,
just thinking how beautiful this world
that you get to explore is,
what do you think is the especially beautiful idea?
The fact that like object level,
what objects are and some notion of objectness emerges
from these models by just like self-supervised learning.
So for example, like one of the things like the dyno paper
that I was a part of at Facebook is,
the object sort of boundaries emerge
from these representations.
So if you have like a dog running in the field,
the boundaries around the dog,
the network is basically able to figure out
what the boundaries of this dog are automatically.
And it was never trained to do that.
It was never trained to,
no one taught it that this is a dog
and these pixels belong to a dog.
It's able to group these things together automatically.
So that's one.
I think in general that entire notion that
this dumb idea that you take like these two crops
of an image and then you say that the features
should be similar,
that has resulted in something like this,
like the model is able to figure out
what the dog pixels are and so on.
That just seems like so surprising.
And I mean, I don't think a lot of us even understand
how that is happening really.
And it's something we are taking for granted,
maybe like a lot in terms of how we're setting up
these algorithms, but it's just,
it's a very beautiful and powerful idea.
So it's really fundamentally telling us something about,
that there is so much signal in the pixels
that we can be super dumb about it,
about how we're setting up the self-supervised learning
problem and despite being like super dumb about it,
we'll actually get very good,
like we'll actually get something that is able to do
very like surprising things.
I wonder if there's other like objectness,
other concepts that can emerge.
I don't know if you follow Francois Chollet,
he had the competition for intelligence
that basically it's kind of like an IQ test
but for machines.
But for an IQ test, you have to have a few concepts
that you wanna apply.
One of them is objectness.
I wonder if those concepts can emerge
through self-supervised learning on billions of images.
I think something like object permanence
can definitely emerge, right?
So that's like a fundamental concept which we have.
Maybe not through images, through video,
but that's another concept that should be emerging from it.
Because it's not something that,
like we don't teach humans that this isn't,
this is like about this concept of object permanence,
it actually emerges.
And the same thing for like animals,
like dogs I think actually permanence automatically
is something that they are born with.
So I think it should emerge from the data.
It should emerge basically very quickly.
I wonder if ideas like symmetry, rotation,
these kinds of things might emerge.
So I think rotation probably, yes, yeah, rotation, yes.
I mean, there's some constraints in the architecture itself.
Right.
But it's interesting if all of them could be,
like counting was another one.
You know, being able to kind of understand
that there's multiple objects of the same kind in the image
and be able to count them.
I wonder if all of that could be,
if constructed correctly, they can emerge.
Because then you can transfer those concepts
to then interpret images at a deeper level.
Right.
Counting I do believe, I mean, should be possible.
You don't know like yet,
but I do think it's not that far in the realm of possibility.
Yeah, that'd be interesting
if using self-supervised learning on images
can then be applied to then solving those kinds of IQ tests,
which seem currently to be kind of impossible.
What idea do you believe might be true
that most people think is not true
or don't agree with you on?
Is there something like that?
So this is going to be a little controversial,
but okay, sure.
I don't believe in simulation,
like actually using simulation to do things very much.
Just to clarify, because this is a podcast
where you talk about, are we living in a simulation often,
you're referring to using simulation to construct worlds
that you then leverage for machine learning.
Right, yeah, for example,
one example would be like to train
an autonomous car driving system.
You basically first build a simulator,
which builds like the environment of the world.
And then you basically have a lot of like,
you train your machine learning system in that.
So I believe it is possible,
but I think it's a really expensive way of doing things.
And at the end of it, you do need the real world.
So I'm not sure.
So maybe for certain settings,
like maybe the payout is so large,
like for autonomous driving,
the payout is so large
that you can actually invest that much money to build it.
But I think as a general sort of principle,
it does not apply to a lot of concepts.
You can't really build simulations of everything,
not only because like one, it's expensive,
because second, it's also not possible for a lot of things.
So in general, like there's a lot of work
on like using synthetic data and like synthetic simulators.
I generally am not very, like I don't believe in that.
So you're saying it's very challenging visually,
like to correctly like simulate the visual,
like the lighting, all those kinds of things.
I mean, all these companies that you have, right?
So like Pixar and like whatever,
all these companies are,
if all this like computer graphics stuff
is really about accurately, a lot of them
is about like accurately trying to figure out
how the lighting is and like how things reflect off
of one another and so on,
and like how sparkly things look and so on.
So it's a very hard problem.
So do we really need to solve that first
to be able to like do computer vision?
Probably not.
And for me, in the context of autonomous driving,
it's very tempting to be able to use simulation, right?
Because it's a safety critical application,
but the other limitation of simulation
that perhaps is a bigger one than the visual limitation
is the behavior of objects.
Because so you're ultimately interested in edge cases.
And the question is,
how well can you generate edge cases in simulation,
especially with human behavior?
I think another problem is like for autonomous driving,
it's a constantly changing world.
So say autonomous driving like in 10 years from now,
like there are lots of autonomous cars,
but there's still going to be humans.
So now there are 50% of the agents,
say which are humans,
50% of the agents that are autonomous,
like car driving agents.
So now the mixture has changed.
So now the kinds of behaviors
that you actually expect from the other agents
or other cars on the road
are actually going to be very different.
And as the proportion of the number of autonomous cars
to humans keeps changing,
this behavior will actually change a lot.
So now if you were to build a simulator
based on just like right now to build them today,
you don't have that many autonomous cars on the road.
So you'll try to like make all of the other agents
in that simulator behave as humans,
but that's not really going to hold true
10, 15, 20, 30 years from now.
Do you think we're living in a simulation?
No.
How hard is it?
This is why I think it's an interesting question.
How hard is it to build a video game,
like virtual reality game,
where it is so real,
forget like ultra realistic
to where you can't tell the difference,
but like it's so nice that you just want to stay there.
You just want to stay there
and you don't want to come back.
Do you think that's doable within our lifetime?
Within our lifetime, probably.
Yeah.
How you tell they are live long.
Ha ha ha ha.
Does that make you sad
that there will be like population of kids
that basically spend 95%, 99% of their time
in a virtual world?
Very, very hard question to answer.
For certain people, it might be something
that they really derive a lot of value out of,
derive a lot of enjoyment and like happiness out of,
and maybe the real world wasn't giving them that,
that's why they did that.
So maybe it is good for certain people.
So ultimately, if it maximizes happiness,
right, I think if it's making people happy,
maybe it's okay.
Again, I think this is a very hard question.
So like you've been a part of a lot of amazing papers.
What advice would you give to somebody
on what it takes to write a good paper?
Grad students writing papers now,
is there common things that you've learned along the way
that you think it takes both for a good idea
and a good paper?
Right, so I think both of these have picked up
from like lots of people I've worked with in the past.
So one of them is picking the right problem
to work on in research is as important
as like finding the solution to it.
So I mean, there are multiple reasons for this.
So one is that there are certain problems
that can actually be solved in a particular time frame.
So now say you want to work on finding the meaning of life.
This is a great problem.
I think most people will agree with that.
But do you believe that your talents
and like the energy that you'll spend on it
will make some kind of meaningful progress
in your lifetime?
If you are optimistic about it, then go ahead.
That's why I started this podcast.
I keep asking people about the meaning of life.
I'm hoping by episode like 220, I'll figure it out.
Not too many episodes to go then.
All right, maybe today, I don't know.
But you're right.
So that seems intractable at the moment.
Right, so I think it's just the fact of
like if you're starting a PhD, for example,
what is one problem that you want to focus on
that you do think is interesting enough
and you will be able to make a reasonable amount
of headway into it,
that you think you'll be doing a PhD for.
So in that kind of a time frame.
So that's one.
Of course there's the second part
which is what excites you genuinely.
So you shouldn't just pick problems
that you are not excited about
because as a grad student or as a researcher,
you really need to be passionate about it
to continue doing that
because there are so many other things
that you could be doing in life.
So you really need to believe in that
to be able to do that for that long.
In terms of papers,
I think the one thing that I've learned
is I've like in the past,
whenever I used to write things
and even now, whenever I do that,
I try to cram in a lot of things into the paper.
Whereas what really matters
is just pushing one simple idea.
That's it.
That's all because that's the paper
is going to be like whatever eight or nine pages.
If you keep cramming in lots of ideas,
it's really hard for the single thing
that you believe in to stand out.
So if you really try to just focus,
like especially in terms of writing,
really try to focus on one particular idea
and articulate it out in multiple different ways,
it's far more valuable to the reader as well.
And basically to the reader, of course,
because they get to,
they know that this particular idea
is associated with this paper.
And also for you because you have,
like when you write about a particular idea
in different ways, you think about it more deeply.
So as a grad student,
I used to always wait toward like maybe in the last week
or whatever to write the paper
because I used to always believe
that doing the experiments
was actually the bigger part of research than writing.
And my advisor always told me
that you should start writing very early on.
And I thought, oh, it doesn't matter.
I don't know what he's talking about.
But I think more and more I realized that's the case.
Like whenever I write something that I'm doing,
I actually think much better about it.
And so if you start writing earlier,
early on you actually, I think get better ideas
or at least you figure out like holes in your theory
or like particular experiments
that you should run to block those holes and so on.
Yeah, I'm continually surprised
how many really good papers throughout history
are quite short and quite simple.
And there's a lesson to that.
Like if you want to dream about writing a paper
that changes the world and you want to go by example,
they're usually simple and that it's not cramming
or it's focusing on one idea and thinking deeply.
And you're right that the writing process itself
reveals the idea.
It challenges you to really think about
what is the idea that explains
that the thread that ties it all together.
And so like a lot of famous researchers
I know actually would start off like,
first they were even before the experiments were in,
a lot of them would actually start
with writing the introduction of the paper
with zero experiments in.
Because that at least helps them figure out
what they're like, what they're trying to solve
and how it fits in like the context of things right now.
And that would really guide their entire research.
So a lot of them would actually first write in intros
with like zero experiments in
and that's how they would start projects.
Some basic questions about people maybe
there are more like beginners in this field.
What's the best programming language to learn
if you're interested in machine learning?
I would say Python just because it's the easiest one to learn.
And also a lot of like programming
in machine learning happens in Python.
So if you don't know any other programming language,
Python is actually going to get you a long way.
Yeah, it seems like sort of a,
it's a toss up question because it seems like Python
is so much dominating the space now.
But I wonder if there's an interesting alternative.
Obviously there's like Swift
and there's a lot of interesting alternatives
popping up even JavaScript.
So I, or are more like for the data science applications
but it seems like Python more and more
is actually being used to teach like introduction
and programming at universities.
So it just combines everything very nicely.
Even harder question.
What are the pros and cons of PyTorch versus TensorFlow?
I see.
You can go with no comment.
So a disclaimer to this is that the last time
I used TensorFlow was probably like four years ago.
And so it was right when it had come out
because so I started on like deep learning in 2014 or so
and the dominant sort of framework for us then
for vision was CAFE which was out of Berkeley.
And we used CAFE a lot, it was really nice.
And then TensorFlow came in
which was basically like Python first.
So CAFE was mainly C++
and it had like very loose kind of Python binding.
So Python wasn't really the first language you would use.
You would really use either MATLAB or C++
to like get stuff done in like CAFE.
And then Python of course became popular
a little bit later.
So TensorFlow was basically around that time.
So 2015, 2016 is when I last used it, it's been a while.
And then what, did you use Torch?
Or did you leave it to PyTorch?
So then I moved to Lua Torch
which was the Torch in Lua.
And then in 2017, I think basically
pretty much to PyTorch completely.
Oh, interesting.
So you went to Lua, cool.
Yeah.
Huh, so you were there before it was cool.
Yeah, I mean, so Lua Torch was really good
because it actually allowed you
to do a lot of different kinds of things.
So which CAFE was very rigid in terms of its structure.
Like you would create a neural network once and that's it.
Whereas if you wanted like very dynamic graphs and so on,
it was very hard to do that.
And Lua Torch was much more friendly
for all of these things.
Okay, so in terms of PyTorch and TensorFlow,
my personal bias is PyTorch
just because I've been using it longer
and I'm more familiar with it.
And also that PyTorch is much easier to debug
is what I find because it's imperative in nature
compared to like TensorFlow, which is not imperative.
But that's telling you a lot that basically
the imperative design is sort of a way
in which a lot of people are taught programming
and that's what actually makes debugging easier for them.
So like I learned programming in C++.
And so for me, imperative way of programming
is more natural.
Do you think it's good to have kind of these two communities,
this kind of competition?
I think PyTorch is kind of more and more becoming dominant
in the research community,
but TensorFlow is still very popular
in the more sort of application machine learning community.
So do you think it's good to have that kind of split
in code bases?
Or so like the benefit there is the competition
challenges the library developers to step up their game.
But the downside is there's these code bases
that are in different libraries.
Right, so I think the downside is there.
I mean, for a lot of research code
that's released in one framework
and if you're using the other one,
it's really hard to like really build on top of it.
But thankfully the open source community
in machine learning is amazing.
So whenever like something pops up in TensorFlow,
you wait a few days and someone who's like super sharp
will actually come and translate that particular code
based into PyTorch and basically have figured
that all those nooks and crannies out.
So the open source community is amazing
and they really like figure out this gap.
So I think in terms of like having these two frameworks
or multiple, I think of course there are different use cases
so there are going to be benefits to using one
or the other framework.
And like you said, I think competition is just healthy
because both of these frameworks keep
or like all of these frameworks really sort of keep learning
from each other and keep incorporating different things
to just make them better and better.
What advice would you have for someone new
to machine learning, you know, maybe just started
or haven't even started but are curious about it
and who want to get in the field?
Don't be afraid to get your hands dirty.
I think that's the main thing.
So if something doesn't work,
like really drill into why things are not working.
Can you elaborate what your hands dirty means?
So for example, like if an algorithm,
if you try to train a network and it's not converging,
whatever, rather than trying to like Google the answer
or trying to do something,
like really spend those like five, eight, 10, 15, 20,
whatever number of hours,
really trying to figure it out yourself.
Cause in that process, you'll actually learn a lot more.
Yeah.
Googling is of course like a good way to solve it
when you need a quick answer.
But I think initially, especially like when you're starting out,
it's much nicer to like figure things out by yourself.
And I just say that from experience
because like when I started out,
there were not a lot of resources.
So we would like in the lab, a lot of us,
like we would look up to senior students
and the senior students were of course busy
and they would be like,
hey, why don't you go figure it out?
Because I just don't have the time
I'm working on my dissertation or whatever,
a final PhD students.
And so then we would sit down
and like just try to figure it out.
And that I think really helped me.
That has really helped me figure a lot of things out.
I think in general, if I were to generalize that,
I feel like persevering through any kind of struggle
on a thing you care about is good.
So you're basically,
you try to make it seem like it's good to spend time debugging,
but really any kind of struggle,
whatever form that takes, it could be just Googling a lot.
Just basically anything just sticking with it
and going through the hard thing,
that could take a form of implementing stuff from scratch.
It could take the form of re-implementing
with different libraries or different programming languages.
It could take a lot of different forms,
but struggle is good for the soul.
So like in Pittsburgh, where I did my PhD,
the thing was it used to snow a lot, right?
And so when it was snowed, you really couldn't do much.
So the thing that a lot of people said was snow
builds character because when it's snowing,
you can't do anything else.
You focus on work.
Do you have advice in general for people
you've already exceptionally successful, you're young,
but do you have advice for young people starting out
in college or maybe in high school?
You know, advice for their career, advice for their life,
how to pave a successful path in career and life.
I would say just be hungry.
Like always be hungry for what you want.
And I think like I've been inspired by a lot of people
who are just like driven and who really like go
for what they want no matter what, like,
you shouldn't want it, you should need it.
So if you need something, you basically go
towards the end to make it work.
How do you know when you come across a thing
that's like you need?
I think there's not going to be any single thing
that you're going to need.
There are going to be different types of things that you need.
But whenever you need something, you just go push for it.
And of course, once you, you may not get it
or you may find that this was not even the thing
that you were looking for, it might be a different thing.
But the point is like you're pushing through things
and that actually brings a lot of skills
and brings a lot of like builds a certain kind of attitude
which will probably help you get the other thing.
Once you figure out what's really the thing that you want.
Yeah, I think a lot of people are,
I've noticed the kind of afraid of that
is because one, it's a fear of commitment.
And two, there's so many amazing things in this world.
You almost don't want to miss out on all the other
amazing things by committing to this one thing.
So I think a lot of it has to do with just allowing yourself
to like notice that thing and just go all the way with it.
Just go all the way with it.
I mean, also like failure, right?
So I know this is like super cheesy
that failure is something that you should be prepared for
and so on, but I do think, I mean,
especially in research, for example,
failure is something that happens almost like,
almost every day is like experiments failing and not working.
And so you really need to be so used to it.
You need to have a thick skin.
But and only basically through, like when you get through it
is when you find the one thing that's actually working.
Like Thomas Edison was like one person like that, right?
So I really, like when I was a kid,
I used to really read about how he found like the filament,
the light bulb filament.
And then he, I think his thing was like,
he tried 990 things that didn't work
or something of the sort.
And then they asked him like, so what did you learn?
Because all of these were failed experiments.
And then he says, oh, these 990 things don't work.
And I know that, did you know that?
I mean, that's really inspiring.
So you spent a few years on this earth
of performing a self-supervised kind of learning process.
Have you figured out the meaning of life yet?
I told you I'm doing this podcast to try to get the answer.
I'm hoping you could tell me.
What do you think the meaning of it all is?
I don't think I figured this out.
No, I have no idea.
Do you think AI will help us figure it out?
Or do you think there's no answer?
The whole point is to keep searching.
I think it's an endless sort of quest for us.
I don't think AI will help us there.
This is like a very hard, hard, hard question,
which so many humans have tried to answer.
Well, that's the interesting thing about difference
between AI and humans.
Humans don't seem to know what the hell they're doing.
And AI is almost always operating
under well-defined objective functions.
And I wonder whether our lack of ability
to define good long-term objective functions
or in introspect, what is the objective function under which
we operate, if that's a feature or a bug?
I would say it's a feature, because then everyone actually
has very different kinds of objective functions
that they're optimizing.
And those objective functions evolve and change dramatically
through their course of their life.
That's actually what makes us interesting, right?
Otherwise, if everyone was doing the exact same thing,
that would be pretty boring.
We do want people with different kinds of perspectives.
Also, people evolve continuously.
That's like, I would say, the biggest
feature of being human.
And then we get to the ones that die,
because they do something stupid.
We get to watch that, see it, and learn from it.
And as a species, we take that lesson
and become better and better, because of all the dumb people
in the world that died doing something wild and beautiful.
Ishan, thank you so much for this incredible conversation.
We did a depth first search through the space of machine
learning, and it was fun and fascinating.
So it's really an honor to meet you,
and it was a really awesome conversation.
Thanks for coming down today and talking with me.
Thanks, Lexi.
I mean, I've listened to you.
I told you it was unreal for me to actually meet you in person,
and I'm so happy to be here.
Thank you.
Thanks, man.
Thanks for listening to this conversation with Ishan Mizra.
And thank you to Anit, The Information, Grammarly,
and Athletic Greens.
Check them out in the description to support this podcast.
And now, let me leave you with some words from Arthur C. Clark.
Any sufficiently advanced technology
is indistinguishable from magic.
Thank you for listening, and hope to see you next time.