How Your Brain Beats AI on the Energy of a Light Bulb - Earl K. Miller

[00:00:00]How your brain beats AI on the energy of a light bulb.

[00:00:05]Earl, please.

[00:00:09]>> [applause] >> Well, that was very inspiring, wasn't it?

[00:00:16]So, what I'm here to talk to you about today is something that really threatens many of our democratic institutions, and that is the great energy demands of AI.

[00:00:28]AI consumes enormous amounts of energy.

[00:00:31]So, here's a projection from a few years ago, 4.4% of the total of total US power output is used by data centers, and the future, future being next year, is up to 12 up to as much as 12% and exponentially beyond that.

[00:00:47]So, data centers for AI already use huge and rapidly growing amounts of electricity, adding tens of millions of tons of carbon each year into the atmosphere. Now, I don't need to tell anybody what a what a threat that is.

[00:01:01]But also, each large data center uses hundreds of thousands to several million gallons of water per day for cooling, straining local supplies. It also causes brownouts in local communities.

[00:01:13]And local communities often have very little say over sighting, tax breaks, really anything to do with these data centers, which erodes trust in government.

[00:01:22]And also, the the infrastructure needed to supply power data centers will naturally concentrate political power into the hands of fewer and fewer individuals.

[00:01:34]So, how do we solve this? It seems like AI is kind of inevitable at this point.

[00:01:37]It's kind of be like holding back the tide. AI is coming. What do we do about this real real threat to our our infrastructure, to our environment, and to our democracy?

[00:01:49]Well, we can take a cue from biology.

[00:01:52]Biology does not work this way.

[00:01:55]Biology does not solve problems by using massive amounts of energy. In fact, biology does the opposite.

[00:02:01]Evolution favors energy efficiency.

[00:02:05]And in fact, our human brain, you know, AI is using nuclear power plants and hydroelectric plants right now to power itself, and AI ain't so smart.

[00:02:14]Where our far more capable brains run on only 20 watts of power, the power of a dim light bulb.

[00:02:22]So, how do the brains do this, and why can't we run AI this way?

[00:02:26]Well, this and related questions is something that's been um engaging our laboratory for the past few years, and I'm going to try to answer this at least at least a hypothesis, and maybe it points to a future where we're not consuming as much energy to power things like AI.

[00:02:40]So, to explain this, let me give you a little brief little tutorial about the about the brain. What is thought? Well, it's the stuff of thought is neurons.

[00:02:48]Neurons connect to each other at junctions, connections called synapses, and they signal each other by brief electrical impulses, which we call spikes.

[00:02:57]And here's a cartoon showing that. You see uh spikes moving down these neural pathways as if it's like a as if your brain is a giant telegraph system where these spikes are like little telegraph impulses moving down the wires to other neurons.

[00:03:13]Now, your brains have a lot of this stuff. The cortex is what we're talking about. The outer layer of your brain is where your brain does a lot of higher-level critical thinking that that we uh associate with the cognition. Your cortex alone contains 20 billion neurons and 10 to the 14th synaptic connections.

[00:03:31]Now, I don't know what number 10 to the 14 is. Let's just call it a gazillion.

[00:03:35]It's a lot.

[00:03:37]So, how does this all produce thought?

[00:03:40]Well, the dominant model from the uh from the 20th century of neuroscience, and one that still holds a lot of sway today, is that your cortex is wired like a giant telegraph system, and these synapses, these connections, form digital-like logic gates.

[00:03:55]And here's how it works. Here's in the bottom on the on the right side of the screen, at the very bottom, you you're I'm going to explain a little bit of how we think vision works. You're looking at this visual scene, this bent paperclip, and way at the back of the brain, where you have your primary visual cortex, where initial analysis goes on, it breaks seems to break down the visual scene into tiny little line segments, these edge detectors.

[00:04:19]So, this this bent paperclip is broken down into these little tiny um line segments, edges.

[00:04:26]And then they combine signals at as neurons converge through synapses on the other neurons, and these signals combine at another stage of processing, you get longer edge detectors.

[00:04:35]Then you keep combining, and the edge detectors come together and they form corner detectors. You keep combining, combining, and eventually you get to neurons and circuits that detect whole objects. This is the way we thought the brain worked in the 20th century, and it still holds a lot of them um sway today.

[00:04:52]Now, the idea is with 20 billion neurons in your cortex, if you combine, combine, combine enough, you eventually get the circuits specialized for peace, love, and understanding.

[00:05:03]And this may be familiar to you, this whole concept, because this is how AI works. These large language models are trained on large data sets, and they adjust their associations, they adjust their connections, to capture statistical patterns in the data. AI says it's based on the brain, it's based on this model of the brain.

[00:05:24]But, it raises a question. Paradigm When paradigms change in science, because the current paradigm can't explain newer and newer observations. And one thing that connect that it cannot explain is how does the brain route signals in this vastly complex network of 20 billion neurons and a zillion synapses? How does the brain do it? The brain's got to control itself, right? Our brains have executive functions.

[00:05:49]We have an inner narrative. We're constantly placing ourselves into the into the current where we are now and interpreting everything through that inner narrative.

[00:05:57]We set goals. We make plans.

[00:05:59]And we're flexible. We react differently and behave differently in different situations.

[00:06:06]In other words, our brains have a mission control system that keeps us on task and directs brain processing.

[00:06:11]Somehow our brains have to self-control all these billions of neurons and connections. And this is in sharp contrast to AI where AI just merely responds input. It doesn't have these goals. It doesn't have plans. It doesn't have an inner narrative. It just reacts to inputs. And you could build a complex system that will react to inputs, but our brains have to do more. They have to control themselves. They have to somehow get control of our own thoughts and all these billions of neurons in your cortex.

[00:06:36]And it seems impossibly complex to build a control system based on this model of connectionism.

[00:06:43]It would require a puppet master that could rapidly reconfigure countless neural connections on the sub-second level.

[00:06:50]And what we have pictured on the left here is a manual telephone switchboard operator captured that concept. You have to somehow the under connectionism some something has to switch around all these connections. And I mention this cuz by the way I I lectured to the graduate students in in recent years graduate students have no idea what a manual telephone They can't believe that we used to make phone calls like this.

[00:07:10]>> [laughter] >> Um but we used to, right? And the brain can't work this way.

[00:07:16]So, what's the solution?

[00:07:18]Well, the solution that many of us are coming to is is a signal in the brain that has been overlooked for a number of years. Brain waves may may play a crucial role.

[00:07:27]Now, what brain waves are are rhythmic fluctuations of the electric fields surrounding neurons. I mentioned neurons got these little spikes, these little impulses like telegraph like Morse code.

[00:07:37]Well, there's an electrical environment around around the neurons. And when they spike, they cause ripples in the surrounding electric fields. So, spiking creates ripples in in this electric pond.

[00:07:47]And ironically, they were the first signal that we ever met electrical signal we ever measured from the brain way back in the in the 1900s.

[00:07:56]And but when the shift when the the mid the mid-20th century when focus shifted to individual neurons and its connectome as a model, brain waves were were dismissed as non-functional, the humming of an engine.

[00:08:08]Now, even back then when I was a graduate student in the 20th century, this really made no sense to me because as you as you see up here on the upper left is EEG recordings from a scalp. You probably have all seen these before.

[00:08:18]Brain waves get squiggly lines on the EEG and they should be squiggly because if they they ever go completely flat, you're either dead or in a coma and you don't want that. Um And this idea that the brain waves are somehow epiphenomenal and non-functional never made any sense to me because as you see on the right here, they they correlate highly with um level of cognition and level of consciousness.

[00:08:42]When you have high frequencies like gamma and above, your brain's really engaged. At beta and alpha, you're more concentrated and more relaxed. Get down to lower frequencies and now you're drifting down into drowsiness. And when you get down to very low frequencies, you're either asleep or under general general anesthesia. So, they correlate We've known for a long time they correlate highly with the state of consciousness.

[00:09:04]Um and our thinking that so we this this We used to think that this is this is or people used to think this is largely epiphenomenal, but our thinking this has evolved considerably in the past a couple of decades.

[00:09:17]We now know that brain waves have a strong influence on neural spiking. So, these little ripples in the pond that I mentioned that when spiking happens when the neurons give these spikes, they ripple the pond. Well, these ripples build upon themselves. They grow larger and larger and the ripples form larger waves that take on a life of their own.

[00:09:34]So, here this cartoon this um GIF here is showing like imagine the spike is a is a buoy, the waves now take over him because the the spike is occurring in this in this in in this context is this electrical ocean of moving waves that has a strong influence over how neurons spike.

[00:09:52]So, the waves then could shape when and where neurons spike via process called a factor coupling. A factor coupling is just pure electrical influences in the brain. And I just mentioned that in case you wanted to go do a Google Scholars Scholar search on this term. There's lots of papers, lots of mounting evidence this place a real central role in brain function.

[00:10:11]And having this electrical influences is kind of like overlay laying powerful radio waves onto telegraph system. Now, when just relying on the sig the Morse code signals through the telegraph system, you have waves that travel across the wires and can modulate impulses in the wire. Okay? And that is a whole other level of control that we hadn't considered until recent years.

[00:10:34]So, here's example of how waves can organize thing. Here's a here's a a crowd doing the wave. You see that the wave is moving from the right to left across the screen. And you know, that that whole crowd is organizing themselves not by individuals all so in coordinating all the all these individuals. The wave itself is organizing the crowd. The crowd can self-organize using using a wave.

[00:10:57]And we know that from, um, many studies on a factor coupling that much like sports fans in a stadium, neural spikes actually follow waves. So, here's an example here on the Y axis is just voltage, a measure of voltage. On the X axis is time. And you can see there there's a one one line is a measure of the electric the brain waves, the electric changes in electric field potentials. And you can see them oscillating at a certain frequency. And here's spikes from an individual neuron.

[00:11:24]And we can see much like the crowd stadium crowd doing the wave, the spikes tend to occur whenever the wave is high and the brain energy is high. Then the wave goes low, the brain energy is low, and the neurons don't spike. So, much like sports fans in the stadium, neuron spiking does follow waves.

[00:11:41]So, why is it useful? Why Why are waves useful? Well, they're there. If they're there, the brain is probably using them.

[00:11:46]Energy doesn't just create stuff um um for fun of it. It Everything in biology uses energy. I assume if it's there, it must be there for a reason. So, what's the reason? What What's the What's the point of having your brain do brain do these oscillating brain waves that seem to control spiking?

[00:12:03]Well, brain waves are useful for controlling your own thoughts, for controlling neurons. You have to control these 20 million neurons in your cortex.

[00:12:11]How do you do it? You do it You can You can either wave functions are very simple and efficient, rather than trying to control the individuals in this crowd one by one by giving them instructions or or having somebody direct everybody in this crowd, you have a simple wave function, and the wave function itself allows the crowd to organize on the level of these groups. So, wave functions are very mathematically simple. They're very computationally tractable, and they tell you with one simple equation what the waves are going to do in the immediate future and at distance.

[00:12:43]All with one simple function, so they're ideal for organizing for the brain to self-organize itself.

[00:12:49]And brains also do the waves. Here's an actual recording from the cortex. And what's shown here is an is an array. Um Each square is a different electrode on the surface of the cortex. The electrodes are separated by 1 mm, so it's like a little, you know, grid on top of the cortex. And the color indicates their their electric voltage recording electric voltage level.

[00:13:10]And as you can see, much like the crowd doing the wave here, the brain's doing the wave. This wave is is turning counterclockwise across the surface of the cortex. Okay? Now, whenever you see organization in the brain, when you see something organized like this in the brain, invariably function follows organization. Your brain organizes things for a reason.

[00:13:30]Now, waves can do more These brain waves can do more than just organize the brain into a simple crowd like doing the wave or into circles. Waves can also our natural way produce complex organization.

[00:13:44]And what are thoughts but complex organization?

[00:13:47]Billions of neurons all organizing themselves in a complex fashion.

[00:13:51]That's and waves that are natural way produces kind of complex level organization.

[00:13:56]For example, this consider this metaphor here. If you put bunch of grains of sand on the speaker, then play musical notes through the speaker, the grains of sand form patterns that match the interference patterns between the waves.

[00:14:09]Now, I'm not suggesting that electric waves move around neurons in your cortex, but think of the sand as the spiking of individual neurons. And the sound waves in fact they're same as electrical waves. Waves are waves, it's the same physics. And these sound waves can actually change create patterns of spiking on on on the surface of the speaker.

[00:14:29]Now, thoughts are organized Our thoughts are organized. I hope our thought I'm I'm trying my best right now. I hope all of our thoughts are organized right now.

[00:14:37]So, if thoughts are organized, these billions of neurons must somehow be organized. Now, imagine the difficulty of organizing the sand or the organizing all the spiking cortex one grain of sand at a time sand grain by grain. It's a computationally impossible.

[00:14:52]But, organization is a natural outcome of the influence of wave patterns. And this is basic physics that evolution could exploit. And that's what evolution does. It uses what's in the room. When nervous systems first start developing are complex enough, they probably naturally begin oscillating producing waves, and that is a natural way for the for the brain to organize itself to take advantage of that.

[00:15:14]And here's a couple of huge bonus features. Electrical waves propagate signals around the brain 5,000 times faster than spiking and synaptic transmission. 5,000 times faster. That sounds extremely useful for me to for a self-organizing system.

[00:15:30]And here's the punchline.

[00:15:32]Wave-based computation is extremely energy efficient. Vastly more energy efficient than digital computation.

[00:15:40]So, to explain that, you can actually use waves to perform computation. So, here's the example of doing addition and subtraction with just two sine waves.

[00:15:49]You have the blue wave um moving left to right, the orange yellowish wave moving from right to left, and they're coming straight at one another. And the dotted line shows what happens when you sum the two waves together.

[00:15:59]Again, voltage is on the x-axis, just space where the waves are traveling on the y- on the x-axis. And as you can see, when the waves perfectly line up, you get that line as a wave twice as big.

[00:16:11]And when they cancel each other out, when they're opposing peaks, they flatten out to to a line.

[00:16:16]You've just done addition addition and subtraction with with just um with just two sine waves. And you can sum anywhere along the space, and all the addition and subtraction and all the variables are all happening everywhere simultaneously.

[00:16:28]Right?

[00:16:29]So, how the waves interact determines the operation. Here are the waves We have two simple sine waves coming together, so we just modeled addition and subtraction.

[00:16:37]All the values that you want to do the computation on are carried in different parts of the wave.

[00:16:42]And then the waves interact, the math is done across all the variables simultaneously. All at the same time.

[00:16:48]And that's a huge advantage, which I'll get to in a moment. And this here is just an example of addition and subtraction one frequency and one dimension. Just one frequency and just just simple 1D sine waves. Well, your brain can do this in three dimensions and across multiple frequencies. And in fact, we already know this is not conjecture that you could do math in the brain with this. We already know you can do this.

[00:17:11]There's a long history of doing um math math math operations using analog computation.

[00:17:17]So, there's a here's a tide predicting machine from 1881 that uses wheels and discs. Wheels and discs are the same thing as electrical oscillating waves.

[00:17:25]Same principles.

[00:17:26]And here, just across the street, Vannevar Bush in 1927 built an analog differential analyzer that used wheels and discs to solve 18 variable differential equations. This is complex math. Okay? All using wheels and discs, which is the same principles of that electrical waves use. Waves are waves.

[00:17:45]Okay?

[00:17:46]And the way you set it up is he set up the wheels to model the differential equation you want to solve.

[00:17:52]And different parts of the wheels encode the variables, just like I mentioned for the sine waves. Then when he turns the crank, the whole differential equation gets solved simultaneously. Parallel computation. And parallel computation is a huge, huge, huge advantage over digital computation.

[00:18:08]Um analog computation is information-rich. With digital, you got to break down the world into ones and zeros and the little bits. That's already highly inefficient.

[00:18:16]But the real efficiency comes from parallel computation. Parallel computation is vastly more efficient than digital computation. You want to solve an 18-variable differential equation using digital computation, you got to solve it one step at a time. With waves, it all gets solved in parallel, simultaneously. And that is a huge energy advantage. And our brains are constantly creating the raw materials, these brain waves. Like I said, evolution uses whatever's in the room.

[00:18:40]Evolution uses what's available. This is a natural substrate that can be used for computation and organizing the brain.

[00:18:47]And also, with another bonus, analog computation can explain some of the mysteries of the brain, like consciousness.

[00:18:53]Right now, consciousness is a unified experience. Right now, you're having this experience of sights, sounds, my voice, whatever thoughts and memories my my what are my words are are engaging in you. You're having a unified experience in consciousness. And every theory of consciousness suggests that the cortex needs to somehow unify itself, get on the same page, all connect so you can have this unified experience. Well, our idea is that these brainwave patterns that are produced by analog computations, when they become large and organized enough, they bind the cortex into this unified state creating consciousness. And if you want to read more about that, this is the paper we just recently wrote.

[00:19:31]So, in sum, >> [clears throat] >> if we're going to build AI, and AI seems kind of inevitable at this point, let's not build one that strains our infrastructure and weakens democratic accountability.

[00:19:41]Let's take a cue from biology and learn how our far more capable brains are also far more energy efficient than today's AI today's AI and try to exploit that.

[00:19:51]And more generally, let's invest in basic science.

[00:19:55]It helps explain mysteries like how brains work and and consciousness, but it also science basic science tackles hard problems in energy, climate, public health, so new technologies and new developments strengthen our society instead of undermining it.

[00:20:10]So, I thank you FOR YOUR ATTENTION.

[00:20:13]>> [applause] >> OH, WE'LL BEGIN Q&A.

[00:20:22]>> YEAH. YEAH. MONROE, thank you for that inspiring talk.

[00:20:25]>> Thank you.

[00:20:25]>> You started out with a gazillion synaptic connections. How many degrees of freedom are there in waves in brain waves?

[00:20:33]>> How many degrees of freedom? It First of all, it depends it depends on the dimensionality, so it depends on on the frequencies and and the and the both and both of the three dimensions.

[00:20:44]>> So, that's exactly what I'm asking about. Do you have a concept that sort of lets us get from the 10 to the 14 to this different space that you're describing?

[00:20:52]>> Well, first of all, all the brain is the brain does is reduce dimensionality. You can't possibly run a brain where there's where there's 10 to the 14 synapses and and 20 billion neurons all working independently. They got to somehow work together.

[00:21:05]>> Yeah. So, your brain's a low-dimensional manifold. It takes the outside world and reduces it to low dimensions. And these dimensions that are that are used in analog computation, we know are complex enough to do to do to do high-level math because we've been doing high-level math on it for for actually for for for centuries. So, I don't have an exact number for you for the dimensionality of the of analog computation, but it is enough to do some pretty high complex mathematics.

[00:21:30]>> Okay. Thank you.

[00:21:33]>> Yeah. Hi, Sam.

[00:21:34]>> It's really lovely. I wonder if connect dots a little bit. I'm sorry for not saying what I'm saying.

[00:21:39]>> Hi, Sam.

[00:21:40]>> Hi.

[00:21:41]Um using this wave concept, do you have thoughts on how to increase the energy efficiency of large language models or how to make them more compact? Because my impression of LLMs is that they do not pay particular attention to the physical configuration of the Do you have any thoughts on I may be wrong about that. I just wonder if you have thoughts on how to use this to to improve that technology.

[00:22:03]>> Well, well, in the end large language models are are the their backbone is digital computation. This is sequential computation. There's a better way to do computation. You could create a Turing machine, a computer, that uses analog analog computation just as easily as you can do it using digital computation.

[00:22:19]It's the same principle. You can build a computer that uses analog computation and it'll be vastly more energy efficient just by the fact that you're using analog instead of the digital alternative. And again, this isn't a conjecture. I'm working with a group based out of UCSB that are building semiconductor chips. And they noticed when they started building these chips that they're that they actually give off these electrical waves. So, they started to exploit these waves to actually do computation in these chips and it increased the the energy efficiency of the chips by an order of magnitude.

[00:22:47]Okay? So, you the the main trick here is to go from digital to a more efficient form of computation. Everything else follows from that.

[00:22:59]>> I have a dumb question.

[00:23:00]But why do you think AI is beating digital and not analog? Is it like lack of imagination or talent or >> I I think digital computing was an engineering convenience. Okay, so back when digital computers first came out, we didn't have the the the the technology to to control electrical waves um as as well as we can now. But at one point digital tech digital computing is kind of a bit of a detour.

[00:23:29]Humans have been using analog computation for astrolabs and stuff like that to navigate since like 1100 BC. So anyways, digital is a detour and was done in a kind of an engineering convenience cuz you couldn't control waves as easily as as as we can now. But now we can and that and people are using and beginning to beginning to use it to make computation more efficient. So I suggest we keep going with that.

[00:23:55]>> So digital hasn't been used that long.

[00:23:59]>> Sorry?

[00:24:00]>> The digital method hasn't been used that long.

[00:24:02]>> Since Well, since about the 1940s, 50s we started digital computing. So yeah, then then there's a centuries of analog computing.

[00:24:10]>> Awesome.

[00:24:12]>> Thank you.

[00:24:13]>> Thank you.

[00:24:15]>> [applause]