Keynote-Surya Ganguli

also for contributing your expertise to this summit. Ladies and gentlemen, I now take this opportunity to invite Professor Surya Ganguly, Professor of AI, Neuroscience and Physics, Stanford University. Professor Ganguly’s research sits at one of the most intellectually fertile intersections in science today. Using the mathematics of physics and the insights of neuroscience to understand how intelligence, biological and artificial intelligence, actually works. His work is helping build the theoretical foundations that practice so urgently needs. Please welcome Professor Surya Ganguly from Stanford University.

Thank you. Great, we got the slides. So we went from a world leader to a VC to now a professor now. So we have a little bit of a change of pace. It’s going to get a little bit more technical. And because I’m a professor, there’s going to be an exam at the end. All right, so pay attention. All right, so I’m going to talk about advancing the science and engineering of intelligence. So, the last decade of AI research has led to stunning advances in the engineering of intelligence, yielding AI systems that stand poised to transform our society. Yet, alarmingly, we understand almost nothing about how they work, and we desperately need to. At the same time, our brain is the product of 500 million years of vertebrate brain evolution, and it is still orders of magnitude better than AI along several axes, and we also need to understand why.

So, I work in a unified science of intelligence across both brains and machines that seeks to both understand biological and artificial intelligence and create more efficient, explainable, and powerful AI. Today, I’ll work on understanding and improving intelligence along three lines. Data efficiency, energy efficiency, and melding brains and machines. First, data efficiency. so um ai is vastly more data hungry than humans we get about 100 million words of language experience ai gets 10 trillion it would take us 240 000 years to read everything that ai read so why is ai so data hungry well well in ai error falls off as a power law it falls off very slowly as a power law with the amount of data this is an example of a famous neural scaling law which captured the imagination of industry and motivated significant societal investments in data collection compute and energy but despite the importance of these neural scaling laws discovered over half a decade ago we lack any scientific theory for why they exist for any modern large language model and why they are so slow just last week we posted the first theory to do so from first principles, we could analytically predict the slope of these neural scaling laws and reconnected their shallow slope to the weak surface statistical structure of natural language itself.

The black line is our theory and the colored lines are experiments in modern LLMs. You can see there’s a good match. But can we make the scaling laws better? We actually can. We actually showed, both in theory and practice, that we can bend the slow power law down to a much faster exponential drop. The key idea is that large random data sets are extremely redundant. If you already have a billion random sentences, it’s unlikely that the next sentence is going to tell you very much that’s new. But what if you could find a non -redundant training set in which each new data point is carefully chosen to tell you something new compared to all the other data points?

We developed theory and algorithms to do just this, and that’s what got us the better exponential. In a completely different line of work, we asked if the process of evolution itself could speed up learning. And we showed it actually can. We evolved robot morphologies, shapes of bodies, from generation to generation. And we showed that successive generations could learn faster. They did so by designing the body to be easier to learn to control. This is an example of something called the morphological Baldwin effect. It’s an effect that has long been conjectured in evolutionary theory, but hard to test in the real world. We demonstrated it for the first time in our simulations. Okay, let’s go on to energy efficiency.

AI is vastly more energy hungry than humans. Our brain only spends 20 watts of power, but modern AI can consume 10 million watts. So why is AI so energy hungry? Well, the fault lies in the choice of digital computation itself, where we use very fast and reliable bit flips at every intermediate step of the computation. Now the laws of thermodynamics demand that every fast and reliable bit flip must consume a lot of energy. Biology chose a very different route. It gets the right answer just in time using the slowest, most unreliable intermediate steps possible. Biology does not rev its engine any more than it needs to. It also co -designs computation and physics much better. For example, it directly uses Maxwell’s equations of electromagnetism to do addition, instead of using complex energy -hungry transistor circuits.

So biology matches its computation directly to the native physics of the universe. So to bridge the vast energy gap between brains and machines, we need to rethink our entire technology stack. . from electrons to algorithms, and optimally match computational dynamics to physical dynamics. For example, given a particular computation, what are the fundamental limits on its speed and accuracy under energy constraints? We recently solved this question for the computation of sensing, which every cell has to do. We found fundamental limits on the lowest achievable error achieved by any chemical computer whatsoever. That’s the red curve. And we also found the family of optimal computers that hug this curve. And we showed, remarkably, that these optimal chemical computers behave a lot like something called G -protein coupled receptors, which hide in every single cell, and they do sensing.

So this yields a connection between what neurons do and what optimal physical sensors would do. Popping up a level, in neuroscience, we can now measure non -neural sensors. We can now measure non -neural sensors. We can now measure non -neural sensors. We can now measure not only neural activity, but also energy consumption in the form of ATP usage. the fundamental chemical fuel that powers all life’s processes. We can do this across the entire fly brain. So by analyzing the couple dynamics of neural computation and energy consumption, we discovered that the brain actually works like a smart energy grid, remarkably. The brain can predict where and when energy will be needed in the future, and it produces just the right amount of energy at just the right time, at just the right location.

So in summary, we still have a lot to learn from evolution in our quest to build more energy -efficient AI, but we don’t have to be limited by evolution. We can go beyond evolution to instantiate neural algorithms in quantum hardware that evolution could not discover. For example, we can replace individual neurons with individual atoms. A neuron in different states of firing correspond to atoms in different excited electronic states. So we can do this with the help of neural networks. We can also replace individual synapses between neurons with photons, quanta of light. Just as synapses allow two neurons to communicate, photons allow electronic states of atoms to communicate through photon emission and absorption. So what can we build with this?

As one example, we could build a Hopfield associative memory network. This is the same network that recently won John Hopfield the Nobel Prize in physics. But this is a quantum version this time that can be built with atoms and photons. And we can show that the quantum dynamics endows the memory with superior capacity, robustness, and recall. We can also go beyond this to build quantum optimizers made entirely out of photons. These photonic computers solve optimization problems in interesting new ways, and we can analyze their energy landscape. So the marriage of neurons… …and neural algorithms with quantum hardware leads to an entirely new field that I like to call quantum neuromorphic computing. okay now returning to the brain the marriage of neuroscience and ai enables a powerful new path forward by melding minds and machines as follows imagine a scenario where we read lots and lots of neural activity from the brain then we use ai to build a model or a digital twin of brain circuits then we can do rapid in silico experiments on the digital twin and use explainable ai to understand how it works but we don’t have to stop there we can control the brain too we can use control theory to learn specific neural patterns that we can write into the digital twin to control it then we can transfer these same neural patterns into the actual brain to write into the brain and control the brain in essence we can learn the language of the brain and then speak directly back to it in its own neural language you So, as one example of this program, we recently developed the world’s most accurate digital twin of the biological retina, and we used explainable AI to understand it.

And in Silico, we could reproduce two decades’ worth of experiments in a matter of days. So this shows a general path forward to dramatically accelerating neuroscience discovery using AI. We also carried out this program in mice, where we were able to use AI to read the mind of a mouse. We could look directly at neural activity in the brain of a mouse, and we could decode what it was seeing at the lower level of resolution that mice can see. This shows that we can learn the native language of the visual brain. But we can go further than that to write to the mind of a mouse. By writing in carefully designed neural activity patterns, we could make the mouse hallucinate a particular percept.

In fact, we could control the mouse brain’s soul. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. We could even tell it to do this. So in essence, we could control what the mouse saw by writing directly into the brain using the native language of its brain itself.

We also applied this to epilepsy. Sorry, we also carried out this program in epilepsy where we built a digital twin of the epileptic brain. Our twin could reproduce actual epileptic seizure dynamics across the entire brain. We then used explainable AI to understand how these seizures were starting. Then we used control theory to be able to control the seizure amplitude in the digital twin. Then we injected these same control signals into the actual brain and controlled seizure amplitude in the actual brain. This shows how to meld brains and machines to control epilepsy. Building on all this, we’re actually creating a new startup called Metamorphic. And we’re going to be using this to control the seizure amplitude in the digital twin.

It will work closely with the Enigma project at Stanford University and together Enigma and Metamorphic. will scale up the construction of digital twins to encompass the entire primate brain, starting with the visual brain. Such scaled -up digital twins offer a powerful path forward to building robust biohybrid AI systems that are taught directly by brain data and to treat brain disease in new AI -driven ways. More generally, the possibilities of melding brains and machines are limitless, both to advance AI and to understand, cure, and augment the brain. To close, what I think we really need is a unified science of intelligence that spans both brains and machines to help us understand both biological and artificial intelligence and create more efficient, explainable, and powerful AI.

Importantly, this pursuit must be done out in the open and shared with the world, and it must be done in a way that is both biological and artificial. It must be done with a long time horizon. This makes academia an ideal place to pursue a science of intelligence, and I believe it’s imperative to expand public investment in the academic study of intelligence, because the academic studies of yesterday laid the strong foundation for today’s AI technology, and it will be the academic studies of today that lay the foundation for tomorrow’s technology, enabling us to go beyond large language models and diffusion models and so forth. Despite the huge and exciting advances happening now increasingly, unfortunately, behind closed doors at companies, I’m extremely excited about what the science of intelligence can achieve out in the open for the public benefit of all.

Thank you.

Thank you so much, Professor Ganguly.