History of Neuroscience and AI: How do brains and AI learn

本站添加: 2026-05-26

[00:00:00]so in the last video we got to the first modern artificial neural network that was built as a in 1958 it was called the perceptron so we're right here but in this video I'm just going to take one step back and talk about why it was it occurred to Rosen bladen's collaborators why they had done this in the first place right why did they think that it was possible to build a Computing machine that was inspired by the way that neurons actually worked so we talked about the neuron Theory um and the this idea of dualism whether not um whether or not the your inner life is actually inside your head inside of the orgon just like every other squishy organ that you have in your body um and really the nail in the coffin of dualism didn't really come until 1943 by this very pivotal paper written by mullik and pits okay so in this very uh influential paper they posited and I'll walk through some assumptions they made that neurons can in fact inact logical operations based on the emerging neurobiology knowledge of things like that they're electrically active and other these things are happening now this is kind of a obvious analogy that we talk about a lot in modern days oh like brains are computers right like this is something that people toss around like it's obvious it was not at all obvious back then in fact some of the more um popular analogies of how brains actually worked are like steam engines and stuff like that and basically people are sort of acknowledging the brains are so complicated that they often use the most sophisticated technology that was available at the time to make an analogy for how um how brains work um and during this decade it became obvious mola compit that if you make a certain set of assumptions about how brains work about how neurons can work they can in fact do logical operations of the kind that we think about in symbolic logic they can have or Gates and and Gates and not Gates okay so they assumed that the activity of the neuron is All or Nothing action potential we talked about in that last video that we have a fixed number of synapses that can be excited in order to excite this neuron so this is the All or Nothing action potential that we talked about how is a summation process and that the activity of inhibitor synapsis absolutely prevents the excitation of auron in given time we can relax this a little bit but this is what they assumed at the time and furthermore we assuming that the structure of the network does not change with time and so we have not learning but those W weights the strength and nature of the synapsis are not changing so this is what they assumed okay and if they assume this about neurons they can basically pin an itive computation by even a small network of neurons onto the formal logic that was uh a couple of decades earlier laid out by Russell and Whitehead in prinkipia Mathematica so if you see here an orgate is what you can do if you have a neuron one and neuron two that both make an equal number of contributions to neuron three if either a one or two is active then neuron 3 becomes active you can have an and gate because here you require at least two synapses right so you have you require one and and two to both be activating three for three to Fire and you can have what's called an not gate right so effectively you can implement this by having a one but a two neuron that is inhibitory so has a negative inhibition and so it's only possible for this to hire if one is firing but not two if they're both firing then three does not fire as well now it turns out that this paper was hugely influential it really allowed the study of neurobiology to be tied to the study of computation and formal logic and Mathematics now the details of all of this as it turns out is not exactly right uh neurons are not logical Gates they're really not okay but in 1943 the paper really got every in thinking including Rosen bladder a few a decade later into thinking oh maybe maybe they're not logical Gates exactly but maybe we can think of them as computational units and we can make computational units that mimic the way the biological neurons are able to compute that analogy there even though the details are wrong was hugely influential and largely correct so what happened here is that just like in the same way we have a representation of um this is like doing art right like you have a representation of something that's real you're ignoring a lot of details but largely The Strokes are there and and and furthermore more importantly we're getting the feeling of these biological neurons that are encapsulated in the artificial neural networks and that's what a lot of inspiration comes from so there's inspiration art and there's inspiration in science and technology as well and I would contend that the the rise of modern artificial neur networks is inspired and it was an inspired uh choice in terms of developing the the the first couple of um iterations of the perceptron so this is the Marin perceptron um back then it wasn't something that lives on your phone in abstraction in code it actually was a jumble of wires it had to be it had to be wired together and Promises were made Okay Rosen plat famously was uh gave a bunch of of um interviews at the time and and to his credit I think he was he had a lot of vision about what these neuron networks could eventually do now what he didn't do was actually deliver on these promises so here's what he told the New York Times back in uh in the late 1950s he promised that perceptrons would be able to recognize people and call out their names and they would be able to hear speech in one language and instantly translate to speech or writing in another language that's what he promised this is like pre1 1960 okay and he all he had were these units and he had a few dozen of these units now you would not be surprised that back in 1960 none of this happened it didn't happen for another 50 years but he had the vision and part of what happened is that it took the next 50 years development for this technology this basic kernel of the idea to get good enough and for the rest of the world to catch up enough that we now actually have those things it has not been that long since we've had AI technologies that are able to to recognize people who call their names and almost instantly hear speech and translate it to another one in writing or in speech this is eventually happened right but he was laughed out of his research funding for many many decades uh because he promised it and he couldn't deliver with the version of the perceptron that he had so we can go here and talk a little bit more about how the perceptron worked because not only was it inspired by the nature of neurons and how neurons worked it had one more key Innovation that made it possible for this to be a good idea to try in the first place I sort of gloss over in this last video this idea of this W here right this W which is a sum of the weights of all of the inputs that gets summed up in this inner product has some kind of bias term and it gets passed through a nonlinear activation function now Mulla and pits like I showed you in The Logical Gates Paper had assumed that these W's are static that nothing is actually changing but what we do know is that these weights in biology also change so around the same couple of decades here we're talking about the 40s and the 50s there arose a theory of learning and activity dependent picity how neurons can actually change the ways they're connected to each other based on what's happening in the past now this eventually became What's called the heavan theory of activity dependent synaptic plasticity there's a lot of different formalisms writing down exactly what it is the cetch version was uh was coined by Carla shatz who famously recapitulated heavan Theory as cells of f together wi are together so when two cells are active at approximately the same time the connection between them the W the the W sub i j that connects cell I to cell J becomes stronger okay so we can write that down um as as a as as some some kind of equation and so Rosen bl's perceptron also had a learning rule so it wasn't just built to work he actually made a uh in inspired directly by HEB HEBs synaptic plasticity equations he wrote down set of equations so that the weights would actually learn from activity and from experience how to make themselves better to accomplish some goal now this learning rule is not the only one you can write down but I'm writing one down for Simplicity here because you can kind of see that the update rule the perceptron used was that the weight of this weight weight um sub I at the next time step is a modification of what it was in the previous time step plus the learning rate and the difference between the true Target and what happened before so here is what's called a prediction error okay so something happened and it wasn't what I wanted but how far off am I is it really far or was it like kind of close so this if it's really close this would be a small number and I would do a small update and if it's really really bad and there's a large discrepancy between what happened and what I wanted to happen then I would multiply that by the Learning Rule and update my weights accordingly okay so this idea and the mathematical formalization of it um is also not only the basis of heavy andity Theory but it's also the basis of reinforcement learning and also uh reinforcement learning in biology and neurobiology how it is that animals and and and living systems learn as well as how we can train artificial and machine systems to learn as well the mathematics is almost exactly the same there's been tons of cross talk over the decades between um the the the the study of the neurobiology of of learning as well as how it is that we can train artificial systems to learn better if they don't start out being able to do something in the way that we wanted to so the synapse is something we've talked about before and one of the things that is really important here is that if we were to update these weights okay we actually need another signal it's not just a matter of uh the W sub I how much we're talking okay and so I told you about at least two types of synapses we have excitatory synapsis and we have inhibitory synapsis now the way this works in biology is that excitatory and um inhibitory neurons use they essentially used a different language they use a different chemical in order to communicate here's an excit excitatory synapse and here's an inhibitory one now common neurotransmitters that are used in M malean brain include acetycholine glutamate Gaba serotonin dopamine these are all probably ones that you've heard of they all slightly different languages they have different chemical properties and importantly it's whatever it has translated by the post synaptic neuro whatever is listening to these messages that then translates into what happens in the down Downstream consequences of having received that message so here are some neurotransmitters that are really common um the aceto choline is the one that's usually what's controlling your muscles so you receive your muscles are listening to acetylcholine when there's acine it contracts okay Gaba is the major inhibitory synapse trans neurotransmitter in the Maman brain this is what's a negative uh a negative synapse is and glutamate is the key major excitatory neurotransmitter in the mamalian brain it's totally different in vertebrate systems but Gaba and glutamate are effectively the negative synapses and the positive synapses that are in the mamalian brain but there's a lot more okay and here's just a couple of them serotonin and dopamine are also active at synapses but they have completely different functions now importantly serotonin and dopamine are really key in providing this learning signal this prediction error this encoding of did I get it right and what you need is a negative signal and a positive signal and also a roar signal from some other neurom modulator to tell the sight of each synapse are you doing it right do you need to get weaker or do you need to get stronger and so that in biology is mediated by a set of neurotransmitters such as dopamine and serotonin there's a couple of other ones that are involved that give you that feedback signal from the external world or whether or not you're getting it right now I just want to show you one more here because um you've seen these chemical organic chemistry structures of these neurotransmitters before they don't have to be neuro they don't have to be little small organic molecules they can in fact be very small proteins um substance P for example is a neuropeptide and it's a transmitter that's responsible for pain sensation in the spinal cord and so if you have um well if you stub your toe and you feel pain then substance P was responsible somewhere in the middle for transducing that information through the neurons so they can get back to your brain and and and and uh and and and andure that something hurt now all of these are key players of neurobiology that then started this whole wave of research over many decades that's been summarized as a some people call it connectionism and it's this idea this philosophy that the function of the network emerges from the connections among them this is a a simple idea that like if you knew the connections among the pieces of the network then surely you can figure out something about how it works that this is at least you know maybe not the sole determining Factor but at least a important factor in terms of determining what the function of the network actually is now where ran into trouble is is uh this is kind of a silly story but but it was this is a true story where the perception ran into trouble is you know Rosen plant went and promise that it would it would you know solve speech translation and facial recognition and essentially go to the Moon right but what happened in the decade sense it only take the next decade is there was a uh there was a a book that was published actually called the perceptron uh by by Minsky and poer where they pointed out that there was actually a fundamental flaw in the perceptron and the flaw was that the single layer perceptron as Rosen blad had built it cannot solve like categorically mathematically cannot solve a class of problems that's exemplified by the xcore problem so I'm explain the xcore problem here it's really simple and it has to do with the fact that like if I'm trying to learn that there's two inputs X1 and X2 okay here they are on the axis and I'm trying to learn the difference between the stars and the Diamonds okay that's all if I ask a perceptron to do it as Ros and blat had built it it actually can't do it right not just it'll be really difficult or it's going to have to try really hard it actually can't solve this problem because it requires a type of non- larity that the perception cannot have now this so upset everyone including Congress that they pulled the budget um and basically a lot of research into artificial neuron networks was there was this huge riseing popularity along with the perceptron and it came to a crashing halt after the publication of this book because it was known to not be able to solve the XR problem so like such a simple problem of course you can solve that problem but this problem cannot be solved by the Petron now in a twist of fate um it turns out that this book not only pointed out this problem and was responsible for the fall of artificial neuron Network research but it also proposed a solution in that if you had multiple layers so you had what's called a multi-layer perceptron instead of a single layer perceptron like the first perceptron was it could solve this problem in fact it's really easy to solve this problem with multi-layer perceptron but that message because it was somewhat political at the time got buried and funding was pulled research stalled and for a really long time the only way that you would be talking about artificial neuron networks in a serious research context was if you were a computational neuroscientist and you were at some obscure conference toly away and still playing with neuron networks even though the world had moved on okay so that happened for a couple of decades and that was really fun and so this is where we're going to end this lecture uh for this video we'll pick up in the next one so what we have here is the is the rise of uh connectionism this idea that neurons can actually compute which inspired the first perceptron okay the first perceptron had activity pendent picity but then it totally crashed and died because it was unable to solve a well-known and relatively simple class of problems such as the XR problem how this resolved and what happened over the next three decades we'll pick up on that in the next video