GABRIEL KREIMAN: We would like to keep this very informal. We have a tradition here where people ask lots and lots of questions. You'll see that many of the speakers have many slides, but this is mostly for you. So the idea here is to have this be a conversation more than lectures. This applies to what I want to say today, but also to all the lectures in the future. So please, please stop and ask questions.

So welcome, everyone, to the Brains, Minds, and Machines summer course. We're very, very excited to have you here. I'd like to introduce here Boris, our co-director. Many of you have met him already. Tommy Poggio is the other course co-director. Unfortunately, he couldn't make it today. He will be here in a couple of days.

And I'd like to start by giving a brief introduction to some of the ideas behind this course. This will be pretty high-level. And then I'll discuss some logistics. And then we'll have another-- I'll give another talk in a few days talking more about specific research and science.

So I'd like to start with a very famous quote by a statistician called IJ Good. And if you haven't read this before, I'll let you read this for a few seconds.

So the idea here is that at some point, we are likely to build machines that are more intelligent than humans. There are physical limits to how fast we can move. For example, we cannot move beyond the speed of light. There are physical limits to temperature. We cannot go below 0 Kelvin.

As far as we know, there is no limit to intelligence. Humans are particularly interesting creatures. They're cute, they're pretty smart, et cetera, et cetera, but there is no reason to think that our intelligence is any kind of supreme or maximum in any kind of distribution.

So in whatever way we define the word intelligence, and there is no consensus on exactly what intelligence is, at some point, I think this will happen. We will have machines that surpass human intelligence.

In many ways, we can argue that this has already happened in very specific domains, some of which may not be super exciting to you. We have machines that can read barcodes in the supermarket much better than humans can. Good luck trying to read those barcodes in the supermarket. We cannot quite do that. So you can call that superhuman intelligence. We have machines that can surpass humans in many tasks along those lines.

And so the claim here is that once we do that, these machines will be so amazing that then that's the last machine we ever need to make because those machines can then build other machines that are smarter and so on and so forth.

And one of the many ways to actually measure-- or one of the many proposed metrics for intelligence is the so-called Turing test that many of you are probably very familiar with. People have been debating for decades now about whether this is a good test or a bad test of intelligence. We can come back to that discussion, but I think it's a very good way of assessing and evaluating what machines can or cannot do in a variety of different tasks.

So in case you haven't heard about this, the Turing test, the basic idea is that you have two rooms. In one of those rooms, you have a machine; in the other one, you have a human. That's the person labeled B in here. And then you have a judge-- in this case, a human judge-- that can ask questions to those two agents. And in those days, those questions were messages that they would pass below the door of that closed room.

And based on those questions and based on those answers, the judge that is Agent C in this diagram would have to be able to identify which room has a human, which room has a machine.

If the judge is unable to determine which room has a human and which one has a machine, then he said this basically-- that means that the Machine A is a very good imitator, that this was also called the Imitation Game. And then in that case, we say that Machine A basically passes the Turing test.

If you haven't read this paper, I strongly recommend that people should read it. It's a fascinating read. Way ahead of its time. And there are lots of fascinating ideas and discussions in there.

In fact, before he introduces this Turing test, he proposes different variations of this, one of which, for example, is a room where you have a man and a woman and you're trying to discriminate which room contains a man, which one contains a woman under situations where they both are trying to deceive you or they're both trying to pretend to be a man or pretending to be a woman. But in any case, so this is one version of how we can try to evaluate machines and their performance.

So you can imagine many different versions of such Turing tests. So for example, we can have a Turing test for barcodes in the supermarket, we can have a Turing test for how much we know about string theory, one for vision, one Turing test for playing chess, for driving, et cetera, et cetera, et cetera. So you can have infinite variations.

So I'll give you a couple of examples, and over the course of the next couple of weeks, we'll give many examples that have to do with vision and visual processing. This is in part because AI has been particularly successful in several aspects of visual processing and because we know a lot about visual processing in brains as well.

So the Turing test for vision would look something like this. We have an image, any arbitrary image-- it's important that we talk about any arbitrary image. And then we can post essentially an infinite number of questions on that image. We can ask, how many people are there? What color are the signs? Are there any dogs? Et cetera, et cetera, et cetera.

So based on these kind of arbitrary questions on arbitrary images, we can ask, again, given the answers to those questions, are those answers coming from a human or are those answers coming from a machine? So there's a lot of important aspects to this kind of testing. One of them, of course, both agents need to be able to understand the question.

So for example, if I ask this question in Arabic or Chinese or Hebrew, if the person doesn't understand the question, that that's not a valid test. Similarly, the computer needs to understand the question. So if the computer doesn't understand the question, it's the same as me posing the question in Esperanto and you don't speak Esperanto. So they need to understand the question.

And then it's important that these questions are arbitrary and infinite in principle. If we're asking only one kind of question, it may be very easy to construct an imitator. The challenge, of course, is to build imitators that can actually surpass or imitate humans in an arbitrary variety of different kinds of tasks.

So Tommy Poggio, who's one of the founding fathers of the Center for Brains, Minds, and Machines, which gave origin to this course; and also, I would argue one of the founding fathers of AI in general, made a claim that if we can understand the brain and understand intelligence, we can find ways to make us smarter and to build smart machines to help us think.

And I like this because I think that it emphasizes the potentially transforming role that understanding intelligence has on almost every aspect of life on Earth. If we really understand intelligence, that may have a major impact in building algorithms to be able to develop large language models as you all know very well.

It will also help us play chess better and develop algorithms that will play chess and Go and so on and build self-driving cars. So there's a lot of engineering applications that many of you are probably quite familiar with and that make the cover of the New York Times quite often.

But in addition to that, one could imagine that understanding intelligence may be able to transform mathematics and physics, and how we interact with each other, and education, curing mental diseases, politics, security. It's hard to think about any aspect of our existence that will not be touched upon or potentially completely transformed if we can really understand intelligence.

So we think that this is not just a question like many other questions. This is really a potentially transforming question that will, in many ways, change history.

OK. Many of you are quite familiar with many different astounding successes of artificial intelligence over the last decade or more.

Here are some examples that many of you are probably very familiar with, from self-driving cars to beating world champions in Jeopardy, chess, or Go. Having systems that you can basically talk to and interact with, all the way to solving problems like trying to determine the three dimensional structure of proteins from the primary amino acid sequence.

I have a very thick accent myself. I was born in Argentina. I'm quite impressed. When I open up a new iPad or phone and just with one or two examples, the machines can understand my accent quite well. In fact, when I talk to people on the street, they often-- I often have to repeat what I'm saying.

Machines can understand me better than when I talk to a random person on the street. So it's quite amazing. Of course, there are many people with Spanish accents, so they have an enormous amount of training.

And then fast forward to today with many of you have probably played with large language models, and we can criticize them in many ways, and we will criticize them in many ways, but it's quite, quite amazing what they can do. It's quite astounding what they can do in terms of their performance.

The problem of protein folding, I confess, I was very, very impressed. There are very, very serious people that have been working on the question of protein folding for decades. When I was a grad student myself back in the pre-history-- my kids like to think that I was basically living at the same time as the Tyrannosaurus Rex and other--

But anyway, so when I was a grad student, I was hesitant about whether I should go into protein folding, and I thought it was a very exciting problem. People have been building detailed biophysical models of interactions and so on for decades.

And in a few years, AlphaFold, a combination of astute algorithms and brute force and a lot of data and computational power, they could beat decades of research into the physics of protein folding, which is one of many examples of when I was wrong in my career.

When Demis Hassabis said that he was going to start working on protein folding, I said, this is not going to work, you're not going to be able to beat decades of research in the physics of protein folding. And here we are, and this is really quite amazing. So this is another example of what I think is a tremendous success of AI.

OK. So back to questions about Turing tests for vision. We'll talk a lot about object recognition. And we've become quite successful at things like-- answering questions like, where are the people in this image? And algorithms in general work quite well. Sometimes they make mistakes.

And just to emphasize what the problem is, I know that many of you are connoisseurs and experts on this, but if you haven't really thought about this problem, this is what the problem looks like. So an image is just a bunch of numbers. So it's a matrix of numbers that denote the intensity of every pixel, perhaps the color of every pixel if you want to work with color images.

From these numbers, we need to be able to infer what that object is. So imagine now, I remove that picture, I just give you those numbers, good luck trying to figure out what that is. But that's somewhat akin to the transformation that happens when light is reflected from objects and impinges on our retina, and then there are retinal ganglion cells that need to send signals to the back of our brain.

So we get some signal that looks more or less like those numbers. From those numbers, we need to be able to infer what's out there. And we'll talk a lot about the algorithms that happen both in the brain, as well as in machines to solve that kind of problem. So loosely speaking, we have a matrix, we want to extract relevant features and use those features for classification.

And a particularly successful way to do that over the last several decades has been to build algorithms that are based on neural networks. So the idea is that we have very simple computational units that are very, very loosely inspired by the idea of neurons in the brain. They are interconnected with each other, hence the term neural network. And those neural networks, when trained appropriately, can do apparently magical things. They can do amazing things.

So there are a lot of emergent computations that happen depending on exactly how those units are connected. And we'll spend a lot of time talking about those connections, those neural networks, what they can do, what they cannot do, how to train them, how to learn from them, how to improve them, and many of the projects in the class will focus on these.

So this is a list of some interesting computations and recipes that have been quite successful in the neural network world, including things like CONVolutional layers, including NORMalization layers, RELU layers, POOL layers, weight changes, dropout, and so on.

And I will argue and I will contend that most of these ideas that have been so prominent and important in the machine learning world. I have actually come from neuroscience. That is from studying the biology of these same problems that need to be solved by actual biological brains.

OK. So here's a semi-random list of a couple of things that neural networks are extremely good at and they can do now, specifically focusing on questions related to vision and pattern recognition. So for many years now, we have algorithms that can recognize handwritten digits, classifying large image data sets like ImageNet.

We have algorithms now that are better at face recognition that what are called superrecognizers or forensic experts. We have algorithms that are better at diagnosing diseases like brain cancer than radiologists. Better than ophthalmologists at diagnosing things like diabetic retinopathy.

And the reason I like this particular story is that we'll have a talk in a few days by people from Google who worked on this project. And what they realized is that they used these kind of images, which is an image of the back of the eye, called a fundus photograph. And they use this image in order to diagnose diabetes of retinopathy, a particular disease of the eye.

But then they realized that from these same images, they could ask different questions, and they asked questions that no clinician had ever thought about before. They could detect the gender of the person, they could detect their age. Moreover, they could detect the risk of cardiovascular disease, which nobody has thought about getting that kind of information from this kind of image. So in a way, they could use machine learning for discovering new principles and new ideas that nobody had thought about before.

We can classify plants, galaxies. If you have a phone, if you have probably played with this, you can recognize plants when you're walking around, recognize galaxies. And then extending to other domains, we have speech recognition, sentiment analysis, decision-making, automatic translation, predictive advertising, earthquakes, protein structure, and so on. So there are many, many astounding successes of AI.

And yet, I will contend that there are many things that deep convolutional networks cannot do, and I want to talk a little bit about some of those things as well.

Before I do that, I want to talk about another angle of AI. So I put this very quickly at the beginning. Can anyone tell what's common to all of these images? What's in common among all of these people? If you have seen me talk about this before or anyone else, don't say it, but--

AUDIENCE: It looks like they're all based on the same face. Some of them are really similar to each other.

GABRIEL KREIMAN: They're very similar to each other. They're based on the same face. What's that?

AUDIENCE: They're not real people.

GABRIEL KREIMAN: They're not real people. OK, good. So I'm preaching to the converted here. You are too smart. So that's good. So you're both right. When I do this with the lay audience, people don't realize, they start to say, well, they're all men. No, that's not true. Oh, they're all colored. No, they're not. They're all smiling. They're not. OK. Anyway. So that's what usually happens, so this doesn't work at all. Anyway, so you're absolutely right. So these are all fake, these people don't exist.

But the reason I wanted to point this out, even though you're both completely right, is that we have amazing generative algorithms now. Not only can we classify lots of things, from faces to breast cancer to galaxies and so on, but we can actually generate things. In this particular case, we can generate images.

And I think there's tons of interesting applications and potentially interesting questions that will come out of the fact that we can actually generate semi-realistic images despite the fact that, in a few seconds, we have at least two people that detected my trick very, very quickly.

OK. So one of the things that-- let's see if this one works, and this is a challenge for everyone, especially for the two of you. So another thing that, of course, has been quite spectacular, especially in the last year or so, is the development of amazing large language models that seem to be quite amazing in terms of their abilities.

So this is an actual Turing test. Actually, this is one of the projects that was started here in the summer course last year led by Mengmi here and with many other people, some of whom are in this room who participated. So this was a conversation between two agents. The two agents are called A and B here. It could be that they're both human, it could be that they're both machines, or it could be that one of them is a machine and the other one is human.

I don't know if the font is large enough. That's yet another reason for people to get cozy and come closer. But I want you to spend a few seconds reading all of these, and then I will ask you whether you think that A is human or not and whether B is human or not. OK. So please read the whole thing.

OK. I'm a very slow reader myself, but hopefully most of you have read this. OK. So raise your hand if you think that A is human. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11. OK, 11. Raise your hand if you think that A is a machine. 1, 2-- OK. I think A, machine 1.

OK. B. Raise your hand if you think that B is human. 1, 2, 3, 4-- OK. Raise your hand if you think that B is a machine. OK. So that one I thought it was 50/50. OK.

So as you can see, it's not that easy. And the fact that I counted, it was exactly 59% for A being machine, and then 50/50 exactly for the second question. It shows that this is not that easy. This is a GPT-3 here in B, and A is a human. So A is a human and B is actually a machine, it's GPT Davinci version for the aficionados. Maybe you're laughing.

AUDIENCE: Oh, I was saying, I will never forget that I got it wrong.

GABRIEL KREIMAN: You got it wrong? OK. So she's the first author and she got it wrong. OK. So this is actually pretty challenging. So if you want to-- just for fun, if you want more examples, you can go to that paper over there.

So we conducted a lot of these experiments. The short answer is that it's becoming pretty challenging to be able to tell. So if you're having a conversation online and talking to some agent, it's becoming not trivial to be able to determine whether you're talking to a human or not in this case. I actually forgot to check exactly what our two vision experts, whether they got it right or wrong. You got both wrong?

AUDIENCE: Both wrong.

GABRIEL KREIMAN: OK, good. OK. So one thing worked. OK, good. All right. There were-- in this case, in that particular version, there were six different tasks. And the question was to try to operationalize the Turing test and to assess to what extent current imitation-- current algorithms can imitate humans or not in a variety of tasks. So this was one of the six tasks, which was a conversation task.

And then there were other tasks, including word association. So I give you a word. I say sky, and I ask you, what's the first word that comes to your mind? And we did that with machines as well. There was image captioning. Similar to the previous one where you have an image and you have to caption that image.

There was one where we had images and you had to detect objects, detect color, and so on. So there were six different tasks. And this is one of infinite possible variations of the Turing test where we're trying to quantify. So if you're interested in Turing-like tasks, come and talk to me or to Mengmi. There's plenty more things to do in here.

So I want to highlight a couple of things that, despite the fact that I've been selling and advocating the normal successes of AI, I'd like to highlight several of the main challenges or several interesting challenges that I think we're still far from solving. So one classical one that has been described now over a decade ago is the notion of adversarial attacks, which many of you are probably quite familiar with.

So you can take an image like that one, for example, which is correctly classified by an algorithm to indicate a label that is a pig. You can introduce a certain amount of noise and convert that to another label-- in this particular case, an airliner. This amount of noise is basically imperceptible, so it's very hard for humans to tell apart these two images.

In fact, you can make the noise so small that you cannot even render it on a computer screen. And yet, you can easily fool most algorithms today in terms of this kind of very basic visual recognition task.

A lot of the enormous successes are from cases where the problems are just too easy or very ill-defined. So this is a classical example where people have claimed about a decade ago that you can actually do action recognition, and the way this is done is by scraping lots of videos from the internet, and then you have things like videos of people playing billiards, cliff-diving, cricket shots, et cetera, et cetera. And then you have-- you write an algorithm to be able to detect whether you have-- whether you can recognize actions or not.

It turns out that this family of tasks is not that challenging. You can actually take a single frame, take pixels from that frame-- nothing's very sophisticated, just pixels-- and be able to do well above chance just by doing that. For example, most of the videos that contain cricket shots have a lot of green, and most of the ones that have writing on board have a lot of black.

So just by detecting the overall dominant color in the image, you can already do quite well, even though that has absolutely nothing to do with action recognition.

And in fact, when you do control tests-- for example, here, the question was, is this person drinking or not? So you build images that are really paired to try to make it really challenging where in one case, the person is drinking, and the other case, the person is making the post as if he were drinking, but he's not actually drinking, this makes the problem very hard. Most algorithms are basically a chance when you actually have adequate controls on the image data set.

OK. So I'm from Argentina, so I cannot give a talk without showing Lionel Messi over there. So the point I want to make here is that this is a little bit outdated, this is 2015, this is 2019. But if you look at the World Cup in robot soccer-- yes, there is such a thing. If you've never thought about this, there is such a thing.

There are people who actually build robots to play soccer. And this is basically what they do. And so if you compare that with-- even I'm not a good soccer player. Even if I were playing soccer, I think I can play better than that.

So I think that there are many problems where I think it's pretty clear that we're extremely far from solving. Part of this is dexterity. So we can debate whether this is intelligence or not. I would contend that there is a lot of intelligence in Lionel Messi and in playing sports in general. And part of it is just the dexterity of being able to have robots that can stand and move and so on, but I think it's pretty clear that the gap here is still quite enormous.

Here's another one that I like very much because it really boils down to basic vision, although I would argue that it actually transcends vision, which is understanding humor in an image. So imagine I give you a picture and I ask you, is that picture funny or not? Is that trying to portray something that's humorous or not?

So this is a very simple binary discrimination. Deliberately we are avoiding text in here, although one could, of course, ask the same question with language as well. And so if I show you one image or another image and I ask you, is it funny or not? Most humans have some intuition about what's funny or not. Humans may disagree. There are cultural influences on humor. Sometimes there are things that are funny to you but not to me, et cetera, et cetera, et cetera.

But all in all, people tend to agree on what they find funny or not. And I would argue that we're still extremely far from being able to have algorithms that can actually understand whether an image is funny or not. And if you're interested in this, we actually have tried this and we have a data sets on humor, and we're trying this sort of thing. There have been some claims out there that machines can do it. I hope I didn't offend you.

AUDIENCE: You now I used to work--

GABRIEL KREIMAN: OK, all right. All right. So anyway, so I think that there have been some claims that you can actually show a picture and then you have some of the very large language models integrated with images that can explain why an image is funny.

I think that most of that is overfitting. I don't think it's true. I don't think you can actually do this and really understand any arbitrary image, whether it's funny or not. And if people disagree, I'm happy to discuss this. And if people are interested in working on this, we have actually images and data sets to work.

So where do we go from here. So we have amazing successes in AI, but at the same time, there's an enormous gap with biological intelligence. And so I think that I'd like to turn to-- it's another one of the main pillars of this course and to our thinking about this family of problems, which has to do with neuroscience.

And I always like to quote Oscar Wilde saying, "The great events of the world take place in the brain. It is in the brain, and the brain only, that the great sins of the world take place also." So despite the enormous computational power that we have, we still have a huge number of tasks that biological brains can solve and machines cannot.

And so I will argue that brains have enabled us to go to the moon, to solve Fermat's theorem, to find antibiotics, to elucidate the structure of DNA and the basis of inheritance, and many, many other problems. And I think we're not quite there. And I would argue that we will need to take inspiration from neuroscience for the next chapter in artificial intelligence.

And so in addition to arguing that neuroscience is critical for constraint and inspiration for AI, we also want to understand neuroscience and brain function because of the enormous toll that it has on our world. So these are just some of the many, many available statistics on the huge toll of mental disease that's prevalent throughout the world.

So we don't need to go into specific numbers. Just to note that mental disease is a major problem in terms of young people, in terms of older people, and basically in terms of it affects everyone. And if we want to fix that, it's not going to be sufficient to be able to build algorithms that can play Go and chess very well. We need to actually go inside the brain and figure it out-- and figure out how the system works.

How do we study brains? So many of you are connoisseurs and are experts. If you're not, people have been developing lots and lots of different techniques to study brains at many different spatial and temporal resolutions. So here is a diagram that on the x-axis, you have the temporal resolution used to study the brain from milliseconds all the way to months. On the y-axis, you have the spatial resolution from studying brains at the level of synapses all the way to whole brains.

And I would argue that there is a spatial resolution, a gold standard to examine brain function, which has to do with the scale of microns and the scale of milliseconds because we know that a lot of the computations in the brain happen at these spatial and temporal scale, and we really need to understand computations at this particular level.

So many of the algorithms that we have constitute only a very poor and very simple approximation to the computations that transcend in biological tissue. So here on the left, you see a staining of an actual neuron, and this whole complexity that you see in there is often approximated by something that looks more or less like this in the world of neural networks.

So in the world of neural networks, we talk about inputs from another neurons, from other neurons. Let's call them presynaptic neurons. We talk about a cell body that linearly integrates the activity from all the inputs. So the x here correspond to the inputs, the w corresponds to the weights, which you can think of as the strength of those synapses. Those are linearly integrated. Then there may be some nonlinearity and activation function, and that neuron produces an output.

So this is at the very heart of, I would say, almost all, if not the vast majority, of neural network models. And even this, I think, it's very clear that people have been discovering and understanding, that there's an enormous complexity even at the most basic level at the level of the computations that happen in a single neuron.

My PhD mentor, Christof Koch, wrote a beautiful book called The Biophysics of Computation. This is an entire book just devoted to that problem. That is, what are the computations that happen in one neuron? What's happening at the level of the dendrites, what's happening at the level of the soma, et cetera, et cetera. So if you're interested in this particular question, I strongly recommend that book. So there's a lot of complexity even at the very bottom of the circuit.

And then perhaps even more relevantly and excitingly, computations in the brain don't happen because of just a single neuron, but it actually takes a village. It's all about the interconnection, it's all about the entire circuitry. So this is what a neural network might look like, and this looks nothing like the type of complexity that we have in the wiring diagrams of actual brains.

So at the architectural level, at the circuit level, there are also massive differences between current neural networks and actual biological brains, and we'll talk a lot about that as well. So I'd like to give a quick shout-out to two animals. OK. I lost my mouse. Anyway. Trust me, there are two more videos on the right. Trust me, I'm a scientist. There were two more videos in there.

Anyway. So animals are amazing. They do a lot of amazing things. They evolved over millions of years to survive and to be able to solve amazing problems.

And the reason I want to bring this up is because I think that there is some Homo sapiens supremacy theory out there that humans are special and that we need to understand human intelligence. I think that there's a lot to learn from animals, and that's why I like to talk about biological intelligence rather than human intelligence.

As excited as I am about humans and interacting with humans, I don't think there is anything really special about humans, we're just one branch of evolution. And in fact, I would contend that most of the progress in neuroscience, most of the progress in terms of understanding circuits and brain function have come from actually studying different types of animal models much more so than from understanding humans and the human brain.

We almost nothing about the human brain, mostly because it's very hard to study human brains. We don't have the tools, we don't have the resolution. There are many things that we cannot do. So I think it's imperative that we need to study animal models if we ever want to make progress on the question of understanding brain function.

So very quickly, this is one seminal example of a profound discovery in neuroscience that has really been extremely influential in neuroscience and vision in general, but also specifically for neural networks. These two gentlemen here are David Hubel and Torsten Wiesel working at the time at the medical school.

And basically, what they did was insert electrodes to listen to the electrical activity of neurons in primary visual cortex, working first in cats, and subsequently in monkeys. And they discovered that there are neurons in primary visual cortex that are tuned to specific features in specific locations of the environment.

So those locations were called the receptive field. Some of those features had to do, for example, with the orientation of a bar. So neurons would fire very strongly for an oriented bar that had 45 degrees, but not for a vertical bar, not for a horizontal bar. They were very extremely sensitive and very specific.

So that incredible degree of specificity was unlike anything that people had discovered before. And for that work, they were awarded the Nobel Prize a few decades ago. Not only that, but they went on to propose a circuitry that were purported to explain some of those features, how that kind of selectivity could come about.

And if you look-- and if you squint a little bit about these diagrams-- this is actually an actual diagram from their papers, what you're seeing here is actual data from the paper, this is a diagram, you can see a very initial diagram reminiscent of the neural networks today.

So many of the neural networks that we have today were inspired by this kind of diagram and this kind of idea of how we can actually create exciting, emergent properties by connecting neurons in the appropriate way.

So fast forward many years. So here's the list of properties that I argued are at the heart of many neural networks today. And I would contend that most of these have an analog from biology. So convolutional layers are aching, to some extent, to filtering operations and the simple cells that David Hubel and Torsten Wiesel discovered.

There's extensive work in neuroscience on the question of normalization, which is very similar to the normalization layers that we have in neural networks. As I said, there are whole books written about the biophysics of computation and individual layers. The input-output curves, what do you get put into the neuron, what do you get out of that neuron? A very, very simplified version of that is the RELU layer.

People have been examining questions about tolerance and invariance in the visual cortex for many years. This is somewhat parallel to the idea of POOL layers. The notion of weight changes is the question of plasticity in neuroscience. People have been studying plasticity in neuroscience both at the synaptic level as well as at the behavioral level for a very long time.

I would contend that the well-known technique of dropout whereby basically some neurons are-- some of the units in the neural network are inactive during some of the epochs during training in order to build tolerance is similar to what happens in the brain in terms of synaptic failures and activity that doesn't propagate from one neuron to the next.

And the idea of deep architectures is very similar to the notion of hierarchical architectures in the brain, which have also been described in neuroscience for quite some time.

OK. I'd like to discuss now very quickly three reasons why I'm optimistic about neuroscience, I'm very excited about neuroscience, and why I think neuroscience has the potential to transform AI as well. Before I do that, I want to get a few-- people have been very silent so far. So a couple of questions, comments. Any thoughts, disagreements? Yes?

AUDIENCE: I'm curious why you thought that neural networks, understanding why certain images are funny was just overfitting, not dimensionally understanding?

GABRIEL KREIMAN: I'd love to see evidence otherwise. I'd like to be able to input any image, particularly those from our data set where we have a pretty good controlled data set, and see what networks do. I've seen examples of very specific images where, for example, the friend has shown me one image that I always like to use in many of my talks, and then the human user asks questions to a large language model about the image. At the end of that--

AUDIENCE: --the large language model know what's in the image? Is it first described to it in words and then it can answer questions about it?

GABRIEL KREIMAN: It's not. So the neural network has the image. There is no description. OK? And then there's a human that asks a question and says, are there people? What are they doing? What's happening? And the human is leading the large language model towards the answer of why the image is funny. And at the end of this dialogue, the large language model says, yes, this is a funny image because blah, blah, blah.

And I may show this on Monday, but I don't think that the computer understood. It's all-- I think all the work was done by the human in leading throughout-- through the questions in that particular case. I haven't--

AUDIENCE: --that the reason correctly for why I thought it was funny is because the human--

GABRIEL KREIMAN: But the human basically walked the model all the way through. what I'd like to see is, if somebody wants to explain-- OK, so I don't know if you found it hilarious or not, but maybe you didn't, but the image of Abraham Lincoln with a phone. Someone, why is that purportedly funny?

AUDIENCE: Because it looks like a bathroom selfie that teenagers do.

GABRIEL KREIMAN: Right. OK. And why is it funny? Why is it funny that Abraham Lincoln--

AUDIENCE: Because--

GABRIEL KREIMAN: OK, right. So OK. So I didn't tell you much. I guided you, now I asked you, why is it funny? But I don't think that-- that image is very famous. Many of you may have seen that image. So it's likely that many models may have been trained with that image. Assuming that the model has not been trained with that image, I'm very skeptical that a model can actually take that image and understand why that image is funny. Or at least I haven't seen any evidence of that. Yes?

AUDIENCE: Although you could argue that you could provide context to the model of saying, well, there's this time period that that occurred. [INAUDIBLE] technological advances and there's-- culture derives how pictures are made. So that there's context there that accumulates to form human. I think the image itself inherently is not really funny, but I think it's the fact that you can incorporate other aspects into it.

GABRIEL KREIMAN: I agree, I agree. And then the question is, why is a human-- I didn't tell him anything, but he could get all of that? But of course, he-- I don't know how old he is. Let's say-- he looks like he's 20. But anyway, he has two decades of experience in this world. So you argue that during those two decades, he got all that context.

So I think that's fair. But I don't-- I'm not disagreeing with you. I think that networks right now don't have those-- the two decades of experience with the world that he has. What do those two decades buy you? They buy you the notion of what a selfie is. They buy the notion that there were no cell phones during the time of Abraham Lincoln.

They buy a lot of things. I think-- and they buy the opportunity to integrate all of that knowledge. I think that we're very far from being able to do that. So I'm not disagreeing. I think we're saying the same thing. You're saying it better, but I think it's the same idea. Yes?

AUDIENCE: I [INAUDIBLE] think that [INAUDIBLE] by giving context without the help of humans. Because like with things that change [INAUDIBLE], and there's a thing of like, let's think step-by-step. And I'm pretty sure if you-- if it starts with us, without the help of a human, it shouldn't be able to infer why it's [INAUDIBLE] understand why step-by-step [INAUDIBLE].

GABRIEL KREIMAN: I accept it. I'm very skeptical. And if anybody-- if anybody is interested, you're welcome to do this as a project for the summer school. We have a data set of images. I'm happy to offer a very, very, very nice price. If you want a dinner-- something of your choice.

I cannot offer Lamborghini, I don't have that kind of money, but I'm happy to offer a dinner in town or whatever to someone that gets a given level of performance on the test data in our control data set. It cannot be overfitting on one of the existing images. I'll give you some training data, we have a test data. You cannot see the test data at all. With the test data, if you get more than 80% performance-- this applies to everyone--

AUDIENCE: Can we do whatever we want?

GABRIEL KREIMAN: You can do whatever you want. You can do training, whatever you want, any model you want. The only thing you cannot do is look at the test data. And that's all. And it has to be our control data set. So for example, here's one silly way that people have published papers. You can say that all the funny images are in black and white and all the non-funny ones are photographs.

Yes, you can write an algorithm with pixels. My daughter, who's in high school, can do that. So you can take pixels and do an SVM and determine which one is funny or not right. That's not really-- it has nothing to do with humor. That's just black and white versus color. So there's lots of confounding factors like that.

But in a controlled data set where you try to get rid as much as possible of those confounding factors, I don't think that that's doable right now. And I'd be happier to be wrong here. So if anybody can prove me that I'm wrong, I'll be very happy. Any other questions? Anything that's not about humor? You want to ask about humor?

AUDIENCE: I have a different question. As you mentioned about [INAUDIBLE] take place at the [INAUDIBLE]. But [INAUDIBLE], for example, do you think AI research can benefit equally from [INAUDIBLE] different scale of neuroscience research?

GABRIEL KREIMAN: No, I think it's not equal. I think the gold standard is to study neurons and circuits of neurons, and I think that other scales are just not as informative for artificial intelligence.

I think that it's going to be very hard to-- I think it can benefit a lot from studying behavior as well, and behavior can be a very important constraint to understand and build better AI, and we'll see many, many examples of that throughout the course. But no, no, I don't think that these different scales are equal. Some are more relevant than others.

AUDIENCE: Are there other fresh stories in the field of AI where AI can infer context?

GABRIEL KREIMAN: Where AI--

AUDIENCE: Can infer context?

GABRIEL KREIMAN: Can infer context. There's a lot of work on trying to infer context. It depends on what you mean by context. Mengmi here has also done a lot of work on using context in visual recognition. For example, the fact that you know you are in this room makes you predisposed to think that there may be chairs or computers, and it's very unlikely that there would be an elephant in here. There's no elephant in the room.