Geoffrey Hinton Unpacks The Forward-Forward Algorithm

seeing a pink elephant notice the words pink and elephant refer to things in the world so what's actually happening is I'd like to tell you what's going on inside my head hi I'm Craig Smith and this is I onai [Music] Jeffrey Hinton a Pioneer in neural networks and the man who coined the term deep learning has been driven throughout his career to understand the brain well his application of the back propagation of error algorithm to deep Networks set off a revolution in artificial intelligence he doesn't believe that it explains how the brain processes information late last year he introduced a new learning algorithm which he calls the forward forward algorithm that he believes is a more plausible model for how the cerebral cortex might learn a lot has been written about the forward forward algorithm in recent weeks but here Jeff gives us a deep dive into the algorithm and the journey that led him to it the conversation is Technical and assumes a lot of knowledge on the part of listeners but my advice for those that don't have that knowledge is to let the technical stuff wash over you and listen instead for Jeff's insights before we begin I'd like to mention our sponsor clearml an open source end-to-end ml op solution you can try it for free at clear.ml that's c-l-e-a-r dot ml tell them I on a I sent you now here's Jeff I hope you find the conversation as fascinating as I did foreign to listeners forward forward networks and why you're looking for something Beyond back propagation despite its tremendous success let me start with explaining why I don't believe the brain is doing back population one thing about back propagation is you need to have a perfect model of the forward system that is impact propagation it's easiest to think about for a layered net but it also works for recurrent Nets for a layered net you do a forward pass where the input comes in at the bottom and goes through these layers so the input might be pixels and what comes out the top might be a classification of is it a cat or a dog you go forwards through the layers and then you look at the error in the output if it says cat when it should say dog that's wrong and you'd like to figure out how to change all the weights in the forward pass so that next time it's more likely to say the right category rather than the wrong one so you have to figure out how a change in a weight would affect the how much it gives the right answer and then you want to go off and change all the weights in proportion to how much they help in getting the right answer and back propagation is a way of figuring out that gradient we're figuring out how much a change in the weight would make the system have less error and then you change the weight in proportion to how much it helps and obviously if it hurts you change it in the opposite direction now about propagation looks like the forward pass but it goes backwards it has to use the same connectivity pattern with the same weights but in the backwards Direction and it has to go backwards through the non-linearity of the neuron there's no evidence that the brain is doing that and there's lots of elements it's not doing that so the worst case is if you're doing back propagation in a recurrent net because then you run the recurrent net forwards in time and it outputs an answer at the end of running forwards in time and then you have to run it backwards through time in order to get all these derivatives so I had to change the weights and that's particularly problematic if for example if you're trying to process video you can't stop and go backwards in time so combined with the fact that there's no evidence the brain does it well no good evidence there's the problem that just for technology it's a mess it interrupts the pipelining of stuff through so you'd really like something like video there's been multiple stages of processing and you'd like to just pipeline the inputs through those multiple stages and just keep pipelining it through and so the idea of the Ford algorithm is that if you can divide the learning the process of getting the gradients you need into two separate phases you can do one of them online and one of them offline and the way you do online can be very simple and will allow you to just pipeline stuff through so the online phase which is meant to correspond to wake you put input into the network and let's take the recurrent version input keeps coming into the network and what you're trying to do for each layer at each time step you're trying to make the layer of high activity or rather high enough activity so that it can figure out that this is real data so the underlying idea is for Real data you want every layer to have high activity and for fake data what comes out we get that later you'd like every layer to have low activity and the task of the network or the thing it's trying to achieve is not to give the correct label as in back propagation is trying to achieve this property but being able to tell the difference between real data and fake data at every layer by each layer having high activity for real data and no activity for fake data so each layer has its own objective function in fact to be more precise we take the sum of the squares of the activities of the units in a layer we subtract off some thresholds and then we feed that to a logistic function that simply decides what's the probability that this is a is real data as opposed to fake data and if the logistical function gets a lot of input it will say it's definitely real data and so there's no need to change anything if it's getting lots of input you won't learn on that example because it's already getting it right and that explains how you can run lots of positive examples without running any negative examples which are fake data because it'll just saturate on positive examples it's getting right so that's what it does in the positive phase it tries to get high sum of squared activities in every layer so that it can tell high enough so it can tell that it's real data in the negative phase which is run Offline that is during sleep the network needs to generate its own data and try and get given its own data as input it wants to have low activity in every layer so the network has to learn a generative model and what it's trying to do is discriminate between real data and fake data produced by its generative model obviously if it can't discriminate at all then what's going to happen is the derivatives that it gets for real data and the derivatives we get for fake data will be equal and opposite so it won't learn anything learning will have finished then if you can't tell the difference between what it generates and real data this is very like again if you know about generative adversarial Networks except that the discriminative net that's trying to tell the difference between real and fake and the generative model that's trying to generate fake data use the same hidden units and so they use the same hidden representations that overcomes a lot of the problems that a gun has on the other hand because it's not doing back propagation to learn the generative model it's harder to learn a good General model that's a rough overview of the algorithm let me ask one a couple of questions on the awake and sleep cycle are you cycling quickly between them okay so most of the research what I would do is the preliminary research cycle quickly between them because that's the obvious thing to do and later on I discovered well I've known for some time that with contraceptive learning you can separate the phases and later on I discovered it worked pretty well to separate the phases recent experiments I've done with predicting characters You can predict you can have it predict about a quarter of a million characters so it's running on real data trying to predict the next character is making predictions he's running with mini batches so after making quite a large number of predictions they're going to updates the weights and then it sees more positive examples it updates away scan so in all those phases it's just trying to get higher activity in the hidden layers but only if it's not already got high activity and you can predict like quarter of a million characters in the positive phase and then switch to the negative phase where the Network's generating its own string of characters and it you're now trying to get low activity in the hidden layers for the characters it's predicting it's looking a little window characters and then you run for quarter of a million characters like that and it doesn't actually have to be the same number anymore we've bought some machines it's very important to have the same number of things in the positive phase and negative phase but with this it isn't the most remarkable is that up to a few hundred thousand predictions it works almost as well if you separate the phases as opposed to interleave and that's quite surprising in human learning certainly in the we can sleep for complicated Concepts that you're learning but there's learning going on all the time that doesn't require a sleep phase well there is in this too if you're just running on positive examples it's changing the weights for all the examples where it's not completely obvious that this is a positive data so it will do a lot of it does a lot of learning in the positive phase but if you go on too long you fails catastrophically and people seem to be the same if I probably sleep for a week you'll go completely psychotic and job hallucinations and you may never recover can you explain I think one thing that people are having trouble non-practitioners are having trouble understanding is the concept of negative data I've seen a few articles where they just put it in quotation marks out of your paper which indicates that they don't understand it okay what I mean by negative data is data that you give to the system when it's running in the negative phase that is when it's trying to get low activity in all the hidden layers and there are many ways of generating negative data in the end you'd like the model itself to generate the negative data so this is just like it was in Baltimore machines the data that the model itself generates is negative data and real data is what you're trying to model and once you've got a really good model the negative data looks just like the real data so no loan takes place but negative data doesn't have to be produced by the model so for example you can train it to do supervised learning by inputting both an image and the label so now the label's part of the input not part of the output and what you're asking it to do is when I input an image with the correct label that's going to be the positive data you want to have high activity you want to input an image with the incorrect label which I just put in by hand that's the incorrect as an incorrect label that's negative data now it works best if you get the model to predict the label and you put in the best of the model's predictions it's not correct because then you're giving it the things it's most the mistake is most likely to make as negative data but you can put in negative data by hand and it works fine and the the reconciliation then at the end is it as in boltzmann machines where you're subtracting the negative data from the positive data but in both machines what you do is you give it positive data real data and you let it settle to equilibrium which you don't have to do with the forward-forward algorithm well not exactly anyway and once I started equilibrium you measure the pairwise statistics that is how often two units that are connected are on together and then in the negative phase you do the same thing with stuff you just let the model settle as producing data itself and you mentioned the same statistics and you take the difference of those pairwise statistics and that is the correct learning signal for a basketball machine but the problem is you have to let the model Settle yeah and there just isn't time for that also you have to have all sorts of other conditions like the connections have to be symmetric there's no evidence Connections in the brain symmetric can you give a concrete example of of positive and negative data in a very simple learning exercise you were working on digits in this example I think is if you're predicting a string of characters the positive data you'd see a little window of characters and should have some hidden letters and because that's a positive window of characters you try and make the activity high in all the hidden layers but also from those hidden layers from the activity in those hidden layers you would try to predict the next character that's a very simple geometry model but notice the geometry model isn't having to learn its own representation so representations are learned just to make positive strings of characters give you high activity in all the hidden notes that's the objective of the learning the objective isn't to predict the next character but having done that learning you've got the right representations for these strings of characters these windows of characters you also learn to predict the next character and that's what you're doing in the positive phase seeing Windows of characters you're changing the weights so that all the hidden layers have high activity for those windows or characters but you're also changing top down weights that are trying to predict the next character from the activity in the hidden lads that's what's sometimes called a linear classifier so that's the positive face in the negative phase you as input you use characters that have been predicted already so you've got this window and you're going along and just predicting the next character and then moving the window along one to include the next character you predicted and to drop off the oldest character you just keep going like that and for each of those frames you try and get low activity in the hidden wires because it's negative data and I think you can see that if your predictions were perfect and you start from a string a real string then the what's happening in the negative phase will be exactly like what's happening in the positive phase right and so the two will cancel out but if there's a difference then you'll be learning to make things more like the positive phase and less like the negative phase and so it'll get better and better at predicting is I understood back propagation on static data there are inputs there's an output and you calculate the error and then you run backwards through the network and correct the weights and then do it again and that's not a good model for the brain because there's no evidence of information flowing backward through the neurons it's not that's not exactly the right wish that there's no no good evidence of derivative information thrown back the studies these error gradients flowing backwards okay obviously the brain has top down connections if you look at the perceptual system there's a kind of forward direction that goes from the thalamus up to him for a temporal cortex where you recognize things and the thalamus is a sort of where the input comes in from the eyes and there's Connections in the backward Direction but the connection in the backward Direction don't look at all like what you'd need for back propagation for example in two cortical areas the connection is coming back don't go to the same cells as connections going forward come from it's not reciprocal in that sense yeah there's a loop between the cortical areas but information in one course got area goes through about six different neurons before it gets back to where it started and so it's a loop it's not uh it's not like a mirrored system okay but my question is you talk about turning the static image into a boring video that allows you to have top-down effects that's right yeah so you have to think of the being a forward Direction which is going from lower layers to higher layers and then orthogonal to that was the time dimension and so if I have a video even if it's a video of just a single thing that stays still I can be going up and down through the layers as I go forwards in time and that's what's allowing you to have top down effects okay I understood that yeah each layer can receive inputs from a higher layer in the previous time step exactly yeah so what a layer is doing it's receiving input from higher layers and lower layers at the previous time step and from itself at the previous time step and if you've got static input that whole process over time looks like a network settling down that's a bit more like a Baltimore machine settling down and the idea is that the time that you're using for that is the same as the time you're using for posting video and because of that if I give you fast input that's changing too fast you can never settle down to interpret it so I discovered this nice phenomenon if you take a new regularly shaped object like a potato for example a nice irregularly shaped potato and you throw it up in the air rotating slowly at one or two revolutions per second you cannot see what shape it is you just can't see the shape of it you don't have time to settle on a 3D interpretation because it's the very same time steps that you're using for posting videos you're using for settling with a static image and what I found fascinating about and maybe this is something that that is already in the literature but this idea of going up and down in the layers As you move through time but it's that's always been in recurrent Nets so to begin with recurrentness we just have one hidden layer so typical lstms and so on would have one hidden there and then Alex Graves the idea of having multiple hidden layers and showed that it was a winner so that idea has been around but it's always been paired with back propagation as the learning algorithm and in that case it was back propagation through time which was completely unrealistic but and the Brain real life is not static so you're not perceiving in a truly static fashion how much of this grew out of Sinclair's contrast of learning or end grads activity differences a couple of years ago I got very excited because I was trying to make a more biologically plausible version of things like Sim clear there's a whole bunch of things like simple it simply wasn't the first of them in fact it's something a bit like simpler that Sue Becker and I published in about 19 1992 in nature but we didn't use negative examples we tried to analytically compute the negative phase and that wasn't there was a mistake it just that would never work um once you start using negative examples then you get things like simply and I discovered that you could separate the phases that they didn't and that got me very excited a few years ago because it seemed like I only had an explanation for what sleep was for one big difference is simply is taking two different Patches from the same image and if they're from the same image it's trying to make them have a similar representation is they're from different images it's trying to make them have different representations sufficiently different once they're different it doesn't try and make them more different and when you think how to say this simply involves looking at two representations and seeing how similar they are and that's one way to measure agreement and in fact if you think about the squared difference between two vectors that decomposes into three terms the sun is to do with the square of the first vector there's something to do with the square of the second vector and then there's the scalar product of the two vectors and the scalar product of the two vectors is the only kind is the only interactive term and so it turns out that squared difference is very like a scalar product a big Square difference is like a small scale of product now there's a different way to measure agreement which is to take the things you'd like to agree and feed them into one set of neurons and now if two sources coming into that set of neurons are green you'll get high activity in those neurons it's like positive interference between light waves and if they disagree you'll get low activity and if you measure agreement just by the activity in a layer of neurons you're measuring an agreement between the inputs then you don't have to have two things you can have as many things as you like you don't have to divide the input into two patches and say to the representation of the two patches agree you can just say I've got a hidden letter does this hidden layer get highly active and it seems to me that's a better way to measure agreement it's easier for the brain to do and it's particularly interesting if you have spiking neurons because what I'm using at present doesn't use Spike Insurance it just says a hidden layer is really asking are my inputs agreeing with each other in which case I'll be highly active or are they disagree in which case I won't but if the inputs arrive at specific times very precise times like spikes do then you can ask not just other stem neurons being stimulated but are they being stimulated at exactly the same time and that's a much sharper way to measure agreement so spiking neurons seem particularly good for measuring agreement which is what I need that's the objective function to get agreement in the positive phase is not in the negative phase and I'm thinking about ways of trying to implant you spiking neurons to make this work better but that's one big difference from simpler that you're not taking two things and saying do they agree you're just taking all the inputs coming into a layer and saying do all those inputs agree when you talk about the activity that's similar to what you were doing with n grads where you're comparing top-down predictions and bottom-up predictions okay okay okay this when you do the recurrent version of the forward algorithm at each time step neurons in a Larry getting top down input and bottom-up input right and they'd like them to agree and if your objective function is to have high activity they'd like to make things highly active there's another version of the forward algorithm where the objective is to have low activity and then you want the top down to cancel out the bottom up and then it looks much more like predictive coding it's not quite the same but it's very similar but let's stick with the version where you're going for high activity you want the top down and bottom up to agree and give you high activity but notice that it's not like the top down is a derivative so in attempts to Implement back crop in neural Nets you try and have top down things which are like derivatives and bottom-up things which are like activities and you try and use temporal differences to give you the derivatives and that's somewhat different here everything's activities you're never propagated derivatives and this algorithm also does away with the idea of dynamic routing that you talked about with yes stacked capsule encoders yeah yes so with capsules I moved on from the dynamic routing to having what are called Universal capsules capsule would be a small collection of neurons and in the original capsules models that collection of neurons would only be able to represent one type of thing like a nose and a different kind of capsule would represent a mouse in Universal capsules what you'd have is that each capsule could represent any type of thing so it would have different activity patterns to represent the different kinds of things that might be there the capsule would be dedicated to a location in the image so a capsule will be representing what kind of thing you have at that location at a particular level of butthole hierarchy so it might be representing you that at the part level you have a nose um and then at a higher level you'd have other capsules that are representing other at the object level you have a face or something but when you get rid of the dedication of a bunch of neurons to a particular type of thing you don't need to do routing anymore and in the forward fold algorithm I'm not doing routine and one of the diagrams in the paper from the product is actually taken from my paper on pothole hierarchies my last paper on capsule models so I had a system called glom an imaginary system and the problem with it was I never had a plausible learning out of it and the thought algorithm is a plausible learning algorithm for glom is something that's neurally reasonable what was fascinating to me at least about capsules is that they captured the 3D nature of reality right lots of neural Nets are now doing that so Nerf models neural Regions Field models now giving you very good 3D models in neural Nets so you can see something from a few different viewpoints and then produce an image of what it would look like from a new viewpoint that's very good for example making smooth videos from frames that are taken a quite long time intervals but in the forward forward algorithm what's your intuition that that this is the if indeed everything works out that this is a model for information processing in the cerebral cortex and that perception of depth and the 3D nature of reality would emerge yeah yeah in particular if I'm showing you a video and the Viewpoint is changing during the video then what you'd want is that the hidden layers should represent 3D structure that's all pie in the sky at present go ahead reach that stage but yeah but with capsules because I think you you referred to pixels having depth so that if one object moved in front of another the system understood that the that it was behind the thing in front of it do you capture that with forward you would want it to learn to deal with that yes yeah I wouldn't wire that in but it's an obvious feature video that it should learn about with babies they learn in just a few days to get structure from motion that is if I take a static scene and I move the Observer or if I take keep the Observer stationary and the experiments were done with a piece of paper folded into a w and if you see it the wrong way around it looks weird and so experiments done by Elizabeth Stokey and other people use the idea that you can tell a lot about the perception of a baby by seeing what they're interested in because they're interested in things that look odd and so they'll pay more attention to things that look hard and within a few days they learn to deal with how 3D structure ought to be related to motion and if you make it related wrong they think it's weird so they learn that very fast whereas it takes them like at least six months I think to learn to do stereo to get it from the true eyes it's just much easier to get from video than from stereo but from evolutionary point of view if something's really easy to learn there's not much Point wiring it in you've been working in Matlab famously now on toy problems are you starting to scale are you still refining I'm doing a bit of scanning I'm using a GPU to make these go a bit faster but I'm still at the stage where there's very basic properties of the algorithm I'm exploring in particular how to generate negative data effectively from the model and until I've got the sort of basic stuff working nicely I think it's silly to scale it up as soon as you scale it up it's slower to investigate changes in the basic algorithm and I'm still at the stage where there's lots and lots of different things I want to investigate for example here's just one little thing that I haven't had time to invest in yet you can use as your objective function to have high activity in the positive phase and low activity in the negative phase and if you do that it'll find nice features in the hidden units or you can have a zero objective function to have low activity in the positive phase if you do that it'll find nice constraints if you think about what physicists do they try and understand nature by finding apparently different things that add up to zero another way of saying is that they're equal and opposite but if you take force and you subtract mass times acceleration you get zero but that's a constraint okay so if you have two sorts of information one of which is force and the other which is mass times acceleration you'd like to have hidden units that see both those inputs and that say zero no activity and then when they see things that don't fit the physics they'll have high activity they'll be the negative data so that's called a constraint and so if you make your objective function B have low activity for real things and high activity for things that aren't real you'll find constraints in the data as opposed to features so features are things that have high variance and constraints of things that have low variance a feature something that's got higher variance and it should have constrained as low various than it should now there's no reason why you shouldn't have two types of neurons one's looking for features and one's looking for constraints and we know with just linear models that a method like principal components analysis looks for the directions in the space at the highest variance they're like features and it's very stable there's other methods like minor components analysis that look for directions in the space that have the lowest variance they're looking for constraints they're less numerically stable but we know that it pays to have both and so that for example is a direction that might make things work better but there's lots there's about 20 things like that I need to investigate and my feeling is until I've got a good recipe for whether you should use features or constraints or both what's the most effective way to generate negative data and so on it's premature to investigate really big systems with regard to really big systems one of the things you talk about is the need for a new kind of computer and I've seen confusion about this too in the Press I've seen people talk about how you talk about getting rid of the annoyman yeah you obviously want computers where the hardware and software are separate yeah and you want them to do things like keep track of your bank account this is for things that where we want computers to be like people to process natural language to process vision all those things that some years ago Bill Gates said computers couldn't do like they're blind and deaf they're not blind and deaf anymore but for processing natural language or doing motor control or doing Common Sense reasoning we probably want a different kind of computer if we want to do a very low energy we need to make much better use of all the properties of the hardware your interest is understanding the brain well I have a side interest in getting low energy computation going and the point about the forward forward is it works when you don't have a good model of the hardware so if for example I take a a neural net and I insert a black box so I have a layer that's just a black box I have no idea how it works it does stochastic things I don't know what's going on the question is can the whole system learn with that black box in there and it has absolutely no problem you've done something different because the black box is changing what happens on the forward pass but the point is it's changing into exactly the same way for both forward passes so it all cancels out whereas in back propagation you're completely sunk at this back box the best you can do is try and learn a differentiable model of the black box and that's not going to be very good if the black box is wandering in its Behavior so the forward algorithm doesn't need to have a perfect model of a forward system it needs to have a good enough model of what one neuron is doing so that it can change the incoming weights of that neuron to make it more active or less active but that's all it needs it doesn't need to be able to invert the forward pass and you're not talking about replacing back propagation which has obviously had enormous success there's plenty of compute plenty of power then back crop is fine but and this is speculative I understand where you are in the research but can you imagine if you had low power computer architecture that that could handle Ford algorithms and you scale them imagine that it would be great I've actually been talking to someone called Jack Kendall who works for a company called rain who is very insightful about what you can do with analog Hardware using properties of the circuits using word for circuits positives of the electrical circuits natural properties of electrical circles um initially that was very interesting for doing a form of Baltimore machine learning but it's also going to be very interesting for the forward algorithm so I can imagine it's scaling up very well but there's a lot of work to be done to make that happen and if it did scale up very well to the degree that large language models have been successful do you think that its abilities would Eclipse those of models based on back propagation I'm not at all sure I think they may not so I think back propagation might be a better algorithm in the sense that a given number of connections you can get more knowledge into those connections using back propagation than you can with the thought algorithm so the Network's moving forward better if they're somewhat bigger than the best size networks for back propagation it's not good at squeezing a lot of information into a few connections back propagation will squeeze lots of information into a few connections if you force it to it is much more happy not having to do that but it'll do it if you force it to and the full algorithm isn't good at that so if you take these large language models so take something with a trillion connections which is about the largest language model that kind of size that's about a cubic centimeter of Cortex and our cortex is like we got a thousand times that much cortex so these large language models that actually know a lot more facts than your ideal because they've read everything on the web not everything but an awful lot yeah the sense in which they know them is a bit dodgy but if you had a sort of general knowledge quiz I think gpg3 even would beat me at a general knowledge quiz there'd be all sorts of people it knows about and when they were born and what they did but I don't know about it and it all fits in a cubic centimeter cortex if you measure by connections so it's got much more knowledge than me I mean much less brain so I think back crop is much better at squeezing information but that's not the brain's main problem broad brains we've got plenty of synapses the question is how do you effectively get information into them how do you make good use of experience David Chalmers talked about the possibility of Consciousness and you're certainly interested in the possibility if you understand how the brain works and you can replicate it this kind of a model let's imagine that it scales beautifully do you see the potential for reasoning and oh I see the potential for reasoning sure but Consciousness is a different kind of question so I think people I'm amazed that anybody thinks they understand what they're talking about when they talk about consciousness they talk about as if we can Define it and it's really a jumble of a whole bunch of different concepts yeah and they're all mixed together into this attempt to explain a really complicated mechanism in terms of an essence yeah so we've seen that before like 100 years ago if you asked philosophers what makes something alive or even if you ask biologists what makes something alive they say Well it has vital force but if you say what is vital force and can we make machines have vital force they can't really Define vital force other than saying is what makes people alive and as soon as you start understanding biochemistry you give up on the notion of vital force you understand about biochemical processes that are stable and things breaking down and so it's not that we cease to have vital force we've got as much vital force as we had before it's just that it's not a useful concept because in an attempt to explain something complicated in terms of some simple essence so another model like that is so sports cars have oomph and some have a lot of them like an Aston Martin with big noisy exhausts and lots of acceleration and bucket seats has lots of and month is an intuitive concept you can ask it doesn't Aston Martin have more umph than my Toyota Corolla and it definitely has a lot more oomph so we really need to find out what oomph is because umph is what's it what it's all about if you're interested in cars or fast cars anyway but the concept of umph it's a perfectly good concept but it doesn't really explain much but if you want to know why is it that when I press the accelerator it goes very fast the concept of oomph isn't going to help you you need to get into the mechanics of it how it actually works and that that's a good analogy because what I was going to say is it doesn't really matter what Consciousness is it matters whether we as humans perceive something as having Consciousness and I think there's a lot to I think there's a lot to be said to that yes yeah so if the if this forward in a large model that scaled relatively low power consumption if it can reason there'll always be philosophers that say yeah but it's not conscious but it doesn't really matter if you can't tell the difference it matters to the philosophers I think it would be nice to show them the way out of their trap they make for themselves which is I think most people have a radical misunderstanding of how terms about perception and experience and sensation and feelings actually work I've had the language works if for example I say I'm seeing a pink elephant notice the words pink and elephant refer to things in the world so what's actually happening is I'd like to tell you what's going on inside my head yeah but telling you what the neurons are doing won't do you much good particularly since all our brains are wired slightly differently it's just no use to you to tell you what the neurons are doing but I can tell you that whatever it is my neurons are doing it's the kind of thing that's normally caused by Pink Elephant being out there if I was doing veridical perception the cause of my brain state would be a pink elephant I can tell you that and that doesn't mean a pink elephant exists in some spooky thing inside my head or it's just a mental thing what it really tells you is I'm giving you a counterfactual I'm saying the world doesn't really contain a pink elephant but if it did contain a pink elephant that would explain my brain stage that plus normal perceptual causation will explain my brain stage so when I say I'm having the experience of a pink elephant the word experience many people think experience refers to some funny internal goings on it's an experience it's some internal no what I'm denoting when I use the word experience is that it's not real I'm took I'm giving you a hypothetical statement but if this hypothetical thing were out there in the world that would explain this brain State and so I'm giving you insight into my brain state by talking about a hypothetical world what's not real about experience is that it's a hypothetical I'm giving them it's not that it lives in some other Spooky World and it's the same for feelings if I say I feel like hitting you what I'm doing is I'm giving you a sense of what's going on in my head via what it would normally cause so in perception it's the world causing of a sexual state with feelings it's the internal State causing an action and I'm giving you insight into my internal state by telling you what kind of action it would cause now I might feel like hitting you or anybody else or kicking the cat or whatever in which case I instead of giving you any one of those actions I just use a term like angry but really that shorthand for all those angry actions so I'm giving you I'm giving you a way of seeing what's going on in my head via describing actions I might do but they're just hypothetical actions and that's what the word feel means when I say I feel typically if I say I feel and then say I feel like blah it's not that there's some special internal Essence that's feeling and computers don't have it computers are just transistors they don't have feeling you have to have a soul to have feeling or something no I'm describing my internal State via the actions it would cause if I were to disinhibit it from another human's point of view if you were a machine and you were saying things like that I would perceive it as you having feelings right so let's take the perception cases it's slightly simpler I think suppose we make a big complicated neural network that can do perception and can also produce language we have those now yeah and so you can show them a minute and they can give you a description what's there and suppose we now take one of those networks and we say I want you to just imagine something and okay so it imagine something and then it tells you what it's imagining so it says I'm experiencing a pink elephant that's experiencing the Pink Elephant just as much as a person is when they say they experience something elephant it's got an internal perceptual state that would normally be caused by a pink elephant but in this case it's not caused by a pink elephant and so it uses the word experience to denote that there you go I think it's got just as much perceptual Sensations as we have although at the current state of large language models don't exhibit that kind of cohesive internal logic you know but they will they will you you think they will oh yeah yeah I don't think I don't think Consciousness is people treat it like it's like the sound barrier that you're either below the speed of sound or you're above the speed of sound you've either got a model that hasn't yet got Consciousness or you've got there it's not like that at all I think a lot of people were impressed by you talking about using Matlab I'm not sure impressed is the right word they were interested they were surprised but what is your day-to-day work like you have other responsibilities but you spend more time on conceptualizing and that could happen while taking a walk or taking a shower or do you spend more time on experimenting like on Matlab or do you spend more time on running large experiments okay it varies a lot over time so I'll often spent a long time like when I wrote that paper about glom I spent a long time thinking about how to organize a perceptual system that was more neurally realistic and could deal with pothole hierarchists without having to do Dynamic setting up and connections and so I spent many months just thinking about how to do that and writing a paper about that I spent a lot of time trying to think about more biologically possible learning algorithms yes and then programming little systems in Matlab and discovering why they don't work so the point about most original ideas is they're wrong and matlab's very convenient for quickly showing that they're wrong and very small toy problems like recognizing handwritten digits I'm very familiar with that task I can very quickly test out an idea to see if it works and I've got I've probably got on my computer thousands of programs that didn't work well that I programmed in an afternoon and an afternoon was sufficient to decide that okay that's not going to work probably that's probably not going to work you never know for sure because there might be some little trick you didn't think of and then there will be periods when I think I've got onto something that does work and I'll spend several weeks programming and running things to see if it works yeah I've been doing that recently with the Ford forward let me see why I use Matlab I learned lots of languages when I was young I learned pop two which was an Edinburgh language UCSD Pascal a lisp common lisp scheme all sorts of lisps and vanilla Matlab which is ugly in some ways but if you're dealing with vectors and matrices it's what you want it makes it convenient and I became fluent in Matlab and I should have learned Python and I should have learned all sorts of other things but when you're old you're much slower learning language and I'd learned plenty of them and I figured since I'm fluent in Matlab and I can test out little ideas in Matlab and then other people can test out running your own big systems I would just stick with testing out things on Matlab there's a lot of things about just literally shaped me but it's also very convenient and you talk a lot about learning in toddlers And is that knowledge base something you accumulated years ago or are you continuing to read and talk to people in different fields I talk to a lot of people and I learned most things from talking to people I'm not very good at reading it I read very slowly and when I come to equations they slow me up a lot so I've learned most of what I know from talking to people and I'm lucky it's only got lots of good people to talk to like I talked to Terry sonoski and he tells me about all sorts of Neuroscience things I talked to Josh Tenenbaum when he tells me about all sorts of cognitive science things I talked to James Howell and he tells me lots of kind of science psychology things so I get most of my knowledge just from talking to people your target nerves you mentioned yeah he corrected my pronunciation of his name look Khan why did you reference him in that talk oh because for many years he was pushing convolutional neural networks oh okay and the vision Community said okay they're fine for little things like handwritten digits but they'll never work for real images and there was a famous paper submitted to a conference where him and his co-workers where he actually did better than any other system on a particular Benchmark I think it was segmenting pedestrians but I'm not quite sure it was something like that and the paper got rejected even though it had the best results and one of the referees so the reason they were rejecting the paper was because the system learned everything so it taught us nothing about vision and this is a wonderful example of a paradigm and the Paradigm for computer vision was you study the task that has to be performed the computation has to be performed you figure out an algorithm that'll do that computation and then you figure out how to implement it efficiently and so the knowledge is all explicit the knowledge that it's using to do the vision is explicit you have to sort it out mathematically and then implement it and sitting there in the program and they just assumed that's the way that computer vision is going to work and because computer vision has to work that way if someone comes along and just learns everything so they're no use to you because they haven't said what the knowledge is what is the heuristic you're using and so it's okay maybe it works but that's just good luck in the end we're bound to work better than that because we're using real knowledge shouldn't we understand what's going on so they completely failed to get the main message which was that it learned everything not quite everything because you're writing convolution but the machine Learning Community they respected him because he's obviously a smart guy but they thought he was on completely the wrong path and they dismissed his work years and years and then when Fife Lee and her collaborators produced the imagenet competition finally we had a big enough data set to show that neural networks would really work well and Jan actually tried to get several different students to make a serious attempt to do the image Network convolutional Nets but he couldn't find a student who was interested in doing it at the same time Elia became very interested in doing it and I was interested in doing it and Alex fishevski was a superb programmer who put a lot of hard work within you into making it work really well so it was very unfortunate for Yan that it wasn't his group that finally convinced the computer vision Community actually this stuff works much better than what you're doing you've now put this paper out there are you hoping to ignite sort of an army of yeah of people trying to put some simple Matlab code out there too yeah because there's a bunch of little things you have to otherwise it won't work and the code needs to get there it's more picky than backup with back propagation you just show people the equations and anybody can go and implement it and it doesn't need a lot of tricks for it to work quite well to work really well it needs lots of Tricks but it's worked quite well it's fine with the forward forwards you need a few tricks for it to work at all the tricks are quite reasonable tricks but once you put them in there then it works and I want to put that Matlab code out there so other people can get it to work but I didn't want to put my very primitive Matlab code out there because it's disgusting [Music] thank you that's it for this week's podcast I want to thank Jeff for his time I also want to thank clear ml for their support we're looking for more sponsors so if you are interested in supporting the podcast please email me at Craig c r a i g at ionai that's e-y-e-hyphen on dot a i as always you can find a transcript of this episode on our website ey hyphen on dot a i I encourage you to read the transcript if you're serious about understanding the forward forward algorithm in the meantime remember The Singularity may not be near but AI is about to change your world so pay attention

Channel: Eye on AI

Views: 134,604

Keywords: deep learning, artificial intelligence, cognitive science, human brain, neural networks

Id: NWqy_b1OvwQ

Length: 58min 55sec (3535 seconds)

Published: Wed Jan 18 2023