Jim DiCarlo: What would it mean to understand intelligence?

nergis, thank you for those inspiring comments, and really thank you for all your support, it's really fabulous, friends, friends,

what

I have to do with you now is the first thing to remind you why we are all here and the nerds already. I did an inspiring job with it. I titled my talk

what

it

would

mean

to

understand

intelligence

, but if there are really only two things you should remember from my talk through this, the first is that

understand

ing human

intelligence

is really the greatest path open. scientific question of all time is up there with the origin of the universe and the origin of life that is in line with what nergis just said this is a long standing question the second thing you should remember is the search for intelligence of the MIT and the center for brains, minds, and machines at the scientific core this is the organization at MIT that directly targets this question it is the only organization that MIT aimed directly at this question so now we're going to say you come back to earth yes This is the goal and we are excited about it, but let's talk about it.

Could we really do it? What

would

it

mean

to understand intelligence? Now this idea of ​​what it would mean has two ways of reading that sentence. What would it mean? In terms of impact, why would we do it and what would constitute a right of understanding? So let's try to break it down a bit. If you just think about what we know about intelligence, maybe we already understand intelligence well. An easy way to ask it is to ask what do we know about current AI, how good is current AI and have there been amazing advances, you are probably reading or hearing about AI every day, but these systems, even with their power, are still being quite limited and the reason we can know this. it's just use using your own eyes your ears you can tell my frustrating interactions with Siri or Alexa the fact that we don't have robots to help us in the kitchen or take care of our loved ones for some reason those don't exist yet why? and maybe what you don't know these are just examples that maybe you don't know that the systems that we have take incredible amounts of data and feed things that are not really sustainable in their current paths, so these are technological problems that reflect the idea that we don't really have an understanding of intelligence and I would like you to quote Feynman what I can't create, I don't understand, in other words, we can't build technologies yet, we must be losing some kind of understanding here now going forward.

On the other hand, you can test more directly what cognitive scientists know about intelligence, what is actually known about human intelligence, and you would think that I could speak to this with some authority after having worked as a department head here for almost a decade of cognitive and brain sciences. that although we have a large amount of knowledge, we have many articles, we have textbooks, we know the regions of the brain involved, we know many of the elementary parts, we really still do not have much understanding of this phenomenon and so on. The reason I can say that is not just because of some authority on the part that is the head of the department, but again I look around to see what it is that we can't impact yet, so if you think about things like the brain disorders, whether they are or not.

If you have a neuropsychiatric or neurodevelopmental disorder, you think of learning difficulties with learning or learning disabilities or even just social conflicts between individuals. These are things that I would suggest to you that if we had a scientific understanding of intelligence, we would have a much greater impact than Currently we can make the quote now that might say here is what I can't fix or I don't know the limits of what I can fix or improve. I still don't understand it in the same sense, we don't have that. scientific understanding, both inabilities to make an impact are really due to an underlying scientific weakness, which is that humans do not yet have a scientific understanding of intelligence and that is what the search is about, it is a community in a search to Now let's understand the intelligence, let's look at that a little bit more when I say all this, well, intelligence, I mean especially human intelligence, of course, we use animal models and others to try to guide our work and understand, I mean in engineering terms and that's important and community, of course.

I'm talking about faculty, students, engineers, scientists, etc., and our followers, okay, let's look at this without human intelligence, why are we talking about human intelligence, as well as brain and cognitive scientists? This is something we know and it is quite surprising. We study that intelligence system every day. study this again in both human and animal models and really the machine behind your eyes that is your brain when coupled with your body can do amazing things navigate in new situations learn with minimal instruction and what others believe use language to communicate writing poetry is about expressing how it feels to collaborate to build bridges and devices and together, collectively, we have built civilizations from nothing.

In short, this is an amazing system and we can do it with very little data and energy, somehow there are things we can still do that for. For example, current AI systems cannot do this, it is simply a remarkable feat. Now let's talk about that. Okay, human intelligence is impressed that we admit that, but what does it mean in engineering terms to understand what that would look like? So I like to start with this quote. Here this is a quote from Francis Crick, your joys, your sorrows, your memories, your ambitions, your sense of personal identity and free will are, in fact, nothing more than the behavior of a vast collection of nerve cells in their molecules. associated, that's what you sometimes call it. the surprising hypothesis is surprising because it is a claim that the mind is something that emerges from the brain that intelligence emerges from the product of a machine now this is surprising to some people it is the working hypothesis of our field of brain and cognitive sciences of that this could be understood as a machine, um, here it is to get to fix ideas by analogy, here is another machine that many of us can have in our pocket or some version of this, um, uh, and the idea here is that There is a similar quote, your favorite extraordinary application. its performance, an amazing user interface, in fact, are nothing more than the behavior of a large set of transistors, um, of course, that is also true, it is a statement that it is a machine, but unlike the machine on the left, the emerging capabilities of how those transistors work. come together and what kind of ways of giving rise to emergent phenomena are understood in engineering terms, but they are not understood on this side, okay, so the result of this what I am trying to tell you is that it is possible to think about that intelligence.

Intelligent human minds are products of machines and can therefore be understood in engineering terms. For example, you could ask how the Mind works in engineering terms, in the same way you could ask how my phone app works in engineering terms, that doesn't mean that your mind or human intelligence will work like your phone, but the idea of ​​the kind of understanding we seek is there. You might ask how perception works, how movement and planning work, how language works, how memory works, these are all questions we ask ourselves. I can ask and answer in engineering terms, so if you think from the phone application, of course the answers don't look like a map from transistors to applications, but of course an understanding of all the intermediate levels that exist and that They allow these transistors to be assembled. to give rise to the emergent property of the notable application now Similarly, if you think about emergent intelligent behaviors as reflected in overt behavior, you can infer the attempt to think about the underlying cognitive processes that are represented in states and algorithmic models of what is happening.

It is revealed as what we call cognition, which is then somehow implemented in brain regions implemented by neural circuits that are ultimately implemented by billions of neurons and trillions of connections, and an ideal scientific level of understanding of engineering would allow us to connect with everyone. These levels doesn't mean that we need to have all of these levels right from day one, but the goal is to make all those connections eventually based on biophysics to say that we have linked the science from one to another, so that's the vision to long term of what this kind of understanding might look like, in this quest to understand intelligence, the applications are things like new possibilities for AI, new ways of teaching our children and ourselves, and new avenues for treating brain disorders and increasing our brain capacity. seeing new ways of intervening that are different than the ways we think today and understanding ourselves more deeply and how we interact with each other and how that might improve our interactions with others, so let's talk about our community here, as I mentioned, it's a broad Community MIT has been working to prepare for this Search.

This community didn't just emerge last week. This is something that's been building, so the community is actually an example of these are faculty hired over the last 10 years or so at the interface of natural and artificial intelligence, many of them are part of the search that you'll hear of many of them today. I, as a Department, was involved in hiring many of these professors and there are many others who are not even listed here, so MIT has been doing significant preparation at the faculty level for this Search. It's really been an effort for the last 10 years to have the center for brains, minds and machines started in 2013 by my colleague Tommy Pojo and Josh Tenenbaum.

You will hear from both. Next, the Center for Brains, Minds, and Machines has allowed us, as an NSF Center, to build an intellectual community again, a broad community of postdoctoral faculty who form the staff and supporters that the community started. Here is a kind of initial version of the cbmm community that has been growing every year. growing more and more and spreading far beyond MIT and that's in part through our training programs and our summer training programs and you'll hear about that later today about how we're trying to grow that community to push this field forward.

We have built teaching and outreach programs, one of which I mentioned first and we have also been training the next generation of students even at the undergraduate level, so as a department head, I was involved in helping start a new career of degree in computing and cognition and this is the number of students at MIT who are in that major and you can see that this is actually one of the fastest growing majors at MIT, which reflects enthusiasm for working on this interface and helps us nurture what the future will require to make these kinds of advances now only organizationally MIT's pursuit of intelligence and its cbmm science course we sit as a research arm within the Schwartzman School of Computer Science the goal of the School of computing is called a faculty and not a school is to integrate many disciplines, especially in this case, the Natural Sciences and Engineering Sciences, but also in its friction with the humanities and social sciences and the School of Administration, for example, and that quest to fulfill this unique role of infusing computing, especially in a natural environment. sciences, but the most important thing is that as we put together these ideas about how human intelligence works, that will change the way Computing happens in the future, so it also feeds back to Computing, we have a provisional space to work in this now and this space is the new version. of this space will exist and I'll show you right here this is the cognitive and brain sciences building that we're actually all sitting in that place this is looking down from above is the new Faculty of Computing building that is It comes here together to us, if you want to see it when you go out to lunch, you can look out this window.

The mission will have space inside this new building with a bridge to the building we are sitting in now, the brain. and kind of the science building, which is the main brain and cognitive science building really in the world and, um, and here's the bridge that went up a few weeks ago, and it's actually more advanced than that, so This is both a physical and intellectual connection between Computing. represented by the university and brain and cognitive sciences, as reflected especially in building 46, but in general throughout theMIT, so we're trying to bring these things together not just philosophically, but in a real way to make real progress the way we do.

That's the strategy we've adopted, so the idea is that we take measurements, discoveries and results from cognitive science from neuroscience and other fields of natural science and try to compare them with theory and model creation and synthesis, to often driven by things like computer science and robotics and bringing them together around what we call computational models or integrated computational models of intelligence these models serve two things simultaneously serve as new hypotheses for the mechanisms of natural intelligence that drive more refined experiments they also serve as new The possibilities of Computing and Engineering again feed back into Computing.

This is the cycle we intend to load and this will also allow these models to be understood at deeper levels through more theoretical analysis to understand their limits and where they can be applied. new forms and the objective of all this, these Integrated Cycles is to build over time a science of intelligence as described in nergis, so why do we believe in that strategy that I just put? Why do we think it can even work well? Because we have seen it. work in the past, so many of us have witnessed firsthand how this strategy has paid off in the area of ​​visual processing, so, for example, you can recognize this person.

Maybe you know first of all that it is a woman's face. You know she can know that she is Marie Curie. You did it very quickly. How did you do it? While it was unclear for many decades how the brand-name condom worked, scientists were busy measuring the behavior of humans and other animals and many of us were hard at work measuring neurons and their connections and the multiple layers within the brain and how they fit together. individual neurons carried. maybe engaging in the calculations to give rise to this surprising behavior of just saying oh, that's a person and I know who she is, that ability was quite a mysterious thing even though we had a lot of data and results that we didn't have. able to tell exactly how that is happening in engineering terms, similarly, computer science and other fields were busy trying to build models that could do this and were struggling for many decades to get this to work as well, somewhere moment the researchers said, hey, let's learn from each other, let's start building models that are inspired by or even very close anatomically to the types of things you see in the brain.

Sciences, let's optimize them with ideas from cognitive science and then amplified by optimization techniques for engineering, this led to new models that are Now, the Intel model, the new intelligent model systems that lead to key advances in both science and engineering, let me explain it on this slide, so I'm just saying that the natural sciences have contributed this type of data and measurements that I referred to together with Engineering, we built these integrated computational models of rapid sensory intelligence. You now know them as deep architectures that were originally inspired by Vision. Now there are deep architectures that are used for deep learning and what people generally call AI nowadays, of course, that's not all AI should be. but that's often what AI is today and it was inspired just by this early sensory processing and by the way, they're also not just technological benefits, they're now the main hypothesis for the mechanisms of the first 200 milliseconds of visual processing .

Perfect hypothesis, in fact the work in progress is to refine these hypotheses more and more, feedback to better models which feed into new engineering, which also helps feedback to better models and both of those things are helping both sides of this engineering thing. , Computing and the Natural. Sciences, we have seen that this strategy works in the Natural Sciences, although it took decades, so the lessons from this, how can we accelerate it if it took decades? What lessons can we take to make it go faster? So lesson one integrates the efforts of Natural Sciences and Computer Engineering, if we do that that produces great benefits in both areas, both disciplines that neither could achieve alone and we do not simply wait for the bottom-up approaches to self-assemble themselves. somehow, being bold with built and integrated testing.

Models have these types of benefits and provide a better understanding of the underlying components that would not have been achieved by studying those components in isolation. We see that in the scientific system of visual assistance where we are studying neurons, we did not know how they worked. We worked individually, we started putting them together into what the hole should do, which gave us a deeper insight into how it really works, we don't learn lesson three, we don't need a perfect understanding of all aspects of the brain components to meaningfully understand intelligence models. What we use today are approximations of what neurons actually do, for example, but they are already making significant movement in the technology space and also making a significant impact in the science space.

Lesson Four, the first 200 milliseconds of visual perception are certainly not all human intelligence and in comparison, we like the models that emerge from them, although they are powerful, we can already see that those models are not capable of doing many of the things that you would like intelligence systems to do and you will hear more about those limits and how we are trying to go beyond them later today. Well, how do we take those lessons and then leverage this strategy to allow teams of people to make big bets on interviews that individual group labs couldn't learn otherwise?

Those lessons, this requires new organizational approaches that we refer to as missions, and significant engineering and personnel resources to enable them. Those enablers are of two types, one is things like benchmarks and platforms for evaluating models to compare natural science data with models, alternative hypotheses. We will hear about that and from Catherine Fairchild we will also design platforms that allow us to build new types of models and extend them. You'll hear about one of our bets on that from Ben Kochman Singa later today, so you'll hear about our big integrated events broadly, but also our Elemental bets that we have to keep feeding that fuel for all the underlying research, so what is the image that you should have in your head from where we are today to the future, which is a science of intelligence, okay, so let's think about that future, so let's take a step back and say, think about what that future would be like .

I want to point out a quote from Patrick Winston, who was the director of the late Patrick Winston, who was the director. of the artificial intelligence lab here at MIT for 25 years and this quote still inspires me today and I hope it inspires many of you, wrote: Imagine a world where human intelligence is truly understood instead of useful but narrow systems like Alexa. and Siri imagine systems as intelligent as us that would change the world now, that sounds cool, but let's think about this a little, change is cool, possibly, but change could also be scary and bad, so the goal of understanding intelligence human has risks.

For example, many of these risks are common to AI technology today, so some of them include things you may already be thinking about or hearing about privacy. Invasion bias. Discrimination and discrimination. Possible job loss. Social or physical weaponry. technologies that result from these types of scientific advances and therefore the way we think about this in the search is that we have at least three ways of thinking about these types of questions, we consider these questions through the lens of our unique scientific approach, which is to try to understand human intelligence, which leads us to think about the pros and cons of these under this type of understanding, so all forms of scientific understanding tend to have pros and cons and this It also has them, for example, the ability to Read Minds may be something that has arisen from a deeper understanding of how human intelligence works and that has a potential risk to society, but it is also what allows Compassionate Care to understand What people need, the ability to perhaps more deeply understand intelligence, has the ability to possibly increase our own intelligence, which in some cases might seem to have negative connotations, but that's the same kind of understanding that could allow us to improve intelligence. learning disabilities again, pros and cons, human jobs could be replaced, especially the things that humans are good at, are the things we could be especially We hope this understanding gives rise, but on the other hand, there are many jobs that people don't really want to do or we need people to do it, relief from unrewarding work as a positive potential of these things, that's how We in the Pursuit think about these things and are aware of these pros and cons as we It advances our research so that we can discuss and guide policy around the potential technological uses of the science of understanding we seek.

Well, then the search is also out of place. a university that allows us to take advantage in this sense of the social and ethical responsibilities of computing from the Faculty of Computing and also our work is fundamentally trying to understand human intelligence, which leads to a strategy that leads us to want to understand human ethics and how it arises, as opposed to the need to incorporate some sort of ethics that might arise from a standard AI approach, okay, just to close things out here. I would like to go back to the quote I said about Patrick, now he said this and I took you. down this road here to say let's think about changing the world, but this still inspires me and why is because I think these things are something we need to pay attention to, but let's think more about the rest of Patrick's quote instead of Knowing what works If you have this understanding instead of knowing what works in K-12, researchers and educators would know why it works.

We could revolutionize the education of people with special needs and provide compassionate care to the elderly and challenged and systems that recognize. How culture influences thinking could help avoid social conflict and mental health could be understood on a deeper level to find better ways to intervene. This again is all from Patrick Winston. I find all of these things still inspiring today for the rest of the day. We'll hear again about many of our ongoing bets and investigations, and my colleague Josh Tenenbaum will explain some of this to you later. What I would like to end with you is to try to imagine a world where human intelligence is truly understood thank you foreigner