Great interview ! ! Small constructive feedback: when Geoff Hinton isn't talking the video shows the "Eye on AI" Logo and (for some reason) that's distracting.
I think that this will open so many possibilities. When working with small MLPs RELU is rearely the best activation function, something like tanh tends to perform much better but if you try to have more than 4 or 5 layers backpropagation chokes on it due to vanishing gradients but with this it wouldn't matter. It doesn't really have to be an exclusive or between Forward-Forward and back propagation, you could train many small backprop networks and join them with the forward-forward algorithm. It won't be as efficient as forward-forward for an analog hardware implementation but it would likely squeeze more into the same amount of weights and will likely provide better accuracy in some tasks. It will also be much less memory demanding than trying to do backprop over the full network and that would increase what our current hardware can do by a lot. Backwards connections would be much more trainable even without the trick of replicating the input data and the layers. With true backwards connections, it may still not converge into a stable solution due to the feedback loop formed, but it won't have the issues of backpropagation through time. If that can be made to work, models can develop something akin to our working memory. Not needing a differentiable model of everything opens the possibilities of inserting stuff in the middle of the network that wouldn't be easy to integrate normally, like database queries based on the output of previous layers or fixed function calculations.
Such a good talk, thank you for organizing this Eye on AI! I have been implementing the FF-algorithm in python and whilst the training is understandable, the testing becomes tricky for multi-class classification trained with the supervised version that Hinton describes. This is because for new examples you don't have any labels, so you need to impose all possible labels on top of the example the same way as in the training and run the network with these to see which has highest hidden layer activation or "goodness" as Hinton describes it. Since the overlayed label is a part of the input, it contributes to the activations, meaning that there is currently no way to test all possible labels at once, which yields to scaling problems for ImageNet or other classification problems with a big amount of possible predictions where every possible class label representation has to be overlayed with the tested input. Will be interesting to see if this can be overcome or if unsupervised learning will be the standard procedure with this technique. Another super-interesting part in my opinion is the fact that Spiking Neural Networks have the Heaviside function as the activation which has no derivative. So traditionally trained SNN's have a Heaviside forward pass and a Sigmoid backwards pass to tune the weights, using FF we will be able to tune SNN's without having to "trick" the backwards pass to not be a step function, which may yield a better representation of our biological processes.
A.I. and WW3 Updates: REPUGNICANS WANT WW3 & CIVIL WAR IN U.S. - AI WILL GIVE IT TO THEM!! Don’t believe it? Ask AI! (We did!) “Commercial Artificial Intelligence” implementations (i.e., Enterprise-wide, mature instantiations) will be very bad for global and local economies, easily replacing all workers, including designers, architects, programmers, analysts, writers, accountants, testing and diagnostics, etc, etc, etc. - ALL (expensive) white collar jobs are the soonest at risk. A.I.-centric CEO’s will MAKE millions being the first to quickly replace all workforce ASAP, starting in the next 12-18 months, as AI “utilities”, then full-blown AI systems and deployments become ubiquitous. It will occur very quickly in the USA. *******Even (especially!) CEO’s will be replaced. Simply put: Using existing corporate data stores and database systems, in the next 12-18 months AI will re-engineer whole economies. Changes will then be implemented, effecting whole market sectors, literally, over night. Only low level, manual labor skills will be highly coveted but, as the global economy crashes, the result will be scaled down work forces everywhere. In the USA, it will become very violent, as ignorant people CONTINUE to lose their jobs with no place to turn for work. Putin’s wartime exit strategy is based on global ollapse to protect his insanity. Xi will sit back and observe, allowing Kim Jong Un to act as a chess board pawn. Kim Jong Un is an angry psychopath, worse than Putin. A.I.: THE WEALTHY ELITISTS’ CRACK PIPE Nearly completed and hoping to keep U. S. distracted, today, the REPUGNICAN’s stinging strategy is more clearly evident, as REPUGNICAN handlers bribe and cajole old and new minions while their elitist controllers are greedily grasping for their newest crack pipe: *******Native mode Artificial Intelligence used to replace the human white collar and blue collar labor forces, as the early robber barons boldly proclaimed and contemplated, aloud. ******* I DARE YOU TO GOOGLE IT! Robber Baron, Jay Gould, repugnantly proclaimed as their elitist goal to control the world and rape Mother Earth to extract her finite resources: **********We will “employ half of America to kill the other half” - Google it.****** We DARE you to seek these (and other) truths! Another greedy psychopath and Gould contemporary, Cornelius Vanderbilt declared, “What do I care about the law. Ain't I got the power?” Google it! And then ask your favorite AI chat bot: Were these well known elitist statements sane or were they the words of a psychopath? Ask soon! Because elitists control all AI technology and future A.I. implementations which are being hacked, and future versions will soon filter (mask) these early conclusions and edit the truth out of and away from their truth-filled responses, as elitists re-program AI bots to omit truth and, instead, invoke the will and desire of REPUGNICAN strategists!
Basically, it is training a neural network but instead of using positive training data, we're using negative training data. This can yield high perplexity due to the fact no one can get "perfect negative data" but we can easily get positive training data; thus I think it will not replace back propagation, but will be very useful in many applications, like neuromorphic hardware; or maybe even applications where we don't even know what the positive data should look like! So we're reverse-solving the problem somehow. This is really very interesting.
I wonder how the forward algorithm, capsules and "GLOM" connect to building those "world models" from observation. I think I understand Yann when he says that you shouldn't make generative models that predict things like pixels, but make predictions about more abstract representations so that you can ignore irrelevant details (like leaves blowing in the trees). Making predictions about higher order, more abstract concepts like "which car overtakes who" etc will make the network start modelling dynamics, and gain an understanding of what it sees, including causal reasoning. Is this Hinton's plan too or does he not think in terms of world models?
this is obvious, real question is how to decide what's relevant and what's not, then this will change with time when system learns new concepts and so generative models have to change, how to make such system stable?
Extremely fascinating to hear this after Chomsky's criticisms of the current deep learning paradigm as failing to differentiate between possible and impossible languages
@@phoneticalballsack Nope. But my lack of personal encounter doesn't seem very important to understanding your statement. Please explain to me, why Chomsky is a dumbass. If you happen to have met him, I'd certainly welcome a anecdote though I don't consider it crucial.
The discussion at 33:32 immediately suggests the possibility of applying a "color" to each neuron, where the squared activation of neurons of one color contributes positively to "goodness", and the squared activation of neurons of the other color contribute negatively to goodness. Any given layer could have neurons of *both* colors. Of course, that leads to additional questions: 1. Is there a rule for determining each neuron's color that could be applied a priori to give better results? 2. Should there be a rule for changing/updating the color of a neuron so the distribution of colors can be adapted to the problem and the data at hand? Finally, to get even farther afield: something whose activation squared counts as positive sounds like a real number. Something whose activation squared counts as negative sounds like an imaginary number. Instead of choosing between one of two colors for neurons, should the activations be multiplied by a *complex* number before squaring, with the sums of the real parts of the squares being used for the objective? Because the effect of complex color is continuous and differentiable, it may be trainable. The network could find, through learning, the balance of importance between features and constraints for the problem domain.
I think the idea of high layer-activations only for the positive data, interesting. The network essentially isn’t giving an Output like in backpropagation, but it’s now the Property of the network to “light up” for correct labels, and therefore indicating whether it’s a positive data or not. I enjoyed this interview given by Hinton about his paper.
25:18 hidden layer is asking: "are my inputs agreeing with each other, in which case I'll be highly active, or are they disagreeing, in which case I won't." :)
Thank you for this interview. Though I don't understand the technical details of it, I did get to draw on some simple things, and also was able to appreciate the serious brain power in Mr. Hinton.
There are other ways to achieve what backprop does, without backprop: use complex, not linear, quantities; use Conversation Theory; use Active Inference. "Attenuation" is a term used by neurosciences for enforcing the "fake data" / "real data" discernment.
Would negative data training be somewhat similar to hypothesis testing? Or at least what they originally conceptualized a null hypothesis as but has now been obscured. Trying to maximize true negatives as opposed to minimizing false positives.
Fascinating model. His view of consciousness doesn't seem as good as Joshua Bach's work though. He says there are a million definitions of consciousness but I believe the most commonly used meaning by philosophers says consciousness is the feeling that its like to be something. Consciousness is a model of a person embedded in a story generated by the neocortex to be stored in memory.
see what Yann says about consciousness in the latest episode: my full theory of consciousness ... is the idea that we have essentially a single world model in our head. Somewhere in our prefrontal cortex and that world model is configurable to the situation we're facing at the moment. And so we are configuring our brain, including our world model for ... satisfying the objective that we currently set for ourselves. ... And so if you have only one world model that needs to be configured for the situation at hand, you need some sort of meta module that configures it, figures out like what situation am I in? What sub goals should I set myself and how should I configure the rest of my brain to solve that problem? And that module would have to be able to observe the state and capabilities - would have to have a model of the rest of itself, of the agent, and that perhaps is something that gives us the illusion of consciousness.
Makes me wonder. Do things like LSD perhaps trigger parts of this 'sleep' state system, but while still awake. Makes quite a bit of sense to me, especially considering how extremely similar 'tripping' hallucinations are to the things AI produces when it is allowed to 'dream away'. Curious.
Oh man the eye was intense! Your all relaxed listening to Hinton's genius and then BAM! a giant spooky eye appears out of nowhere and scares the bejesus outa ya. Intense experience.
I have been trying to find the podcast where Hinton basically says that the longer length of tokens contributes to hallucinations and variance based on standard ML/DL, anybody out there that heard the same thing?
44:56. I think it depends on what is meant by "but it doesn't really matter if you can't tell the difference." Do we simply mean, as long as the illusion is convincing? Like a Hollywood special effect? Or do we mean, it's not "possible" to tell the difference, because it's beyond our capacity to interrogate? The former is a matter of laziness, where we are willing to accept the "optical illusion" because we don't want to understand the magic. Whereas the latter, the situation has moved to a point where we've pushed the investigation to a sort of "event horizon" from which we are bounded from making any further inquiry. I think it very much matters which of these situations we find ourselves in; ethically, if nothing else.
Hi Jeff. As infant animal learners, we output a behavior and get almost immediate feedback from a parent on whether that behavioral output of a moment ago was "good" or "bad." Did mom look away or smile and interact more? This seems like a crude but fair example of back propagation. No? What do you think Mr. Hinton?
Let me see if I understand He is redesigning the black box. Classical black box has explanatory features in the entry and labels or variables to be predicted in the output. In this approach, everything is in the input. And the output is the "hint of simultaneity" of blocks of entries. If that's like so, I would like to stress that this concept is the foundation of all this. The learning algo depends on this structure. One more thought. "Idea association" works this way. "Perception-action" must work in another way. Action looks like an output. Or can it match a FF framework
I’m wondering if the brain isn’t using both the positive and negative training at the same time. Much of daily brain operation is on the negative training. Surprise generates activity. Otherwise not active.
The largest neural network has a trillion connections, which is about a cubic centimeter of the human cortex, which is about 1,000x larger... What a magnificent thing the human brain is!
@@lucamatteobarbieri2493 and for the same amount of computation as the human brain does, uses many times as much energy. Not a problem for a stationary computer, it'd never work for biological beings even if they were born fully formed for their brains and their sizes.
16:34 example of negative data in a negative phase you use characters that have been predicted already.. you're trying to get low activity cuz it's negative data..
54:08 Yann LeCun's convolutional neural networks are fine for little things like handwritten digits but they'll never work for real images says the vision community
When there finally was a big enough data set to show that neural networks would really work well, Yann wanted to take a bunch of different students to make a serious attempt to do the image convolutional neural network work, but he couldn't find a student who'd be interested in doing that :( and at the same time Ilya Sutskever and Alex Krizhevsky, who's a superb programmer, started to be interested in doing that and put a lot of hard work into making it work eventually.. so Yann LeCun deserves to be mentioned, too, according to Geoffrey Hinton
The problem with forward propagation is that it may change its mind to a projection already made and switch fast back again to earlier prediction. However, it is still the better than back propagation. Actually "funny", because negative data is how you get rid of all the BS you don't want to know 🙂
Entirely (as whole) the world data is composed by; good, bad, and hallucinating (half good+half bad) data. You can't make non hallucinating AI with current data. Probable far far in future AI will can solve somehow to be non hallucinating. Or you can make one special AI to filter out the hallucinating data, but is not good idea, lot things to work need hallucinating data. My noob opinion.
Don’t appreciate your eye flickering eye motif repetition gaining frequency in a disturbing way; instead of staying on the guest… you are open to subliminally training … regardless of intent, IS illegal as well as against u tube regulations. Not good either way. Ive documented. Desist.
Geoff chose the wrong acronym. Pink elephant. The N Vietnamese had pink elephants. They rolled in the red clay and became pink. Geoff seems to be taking of absurdity rather than reality. To me pink elephants really are a thing in reality.
Very enlightening video, There’s this woman I got in touch with during the 2020 lockdown which cost me my job. Ms. Norman Davis helped me manage my assets by introducing my to the best trading platform and strategies, I earned a lot of $$$ working with Norman at the comfort of my home. I still keep in touch with the amazing lady
Hi there, I’m commenting from Switzerland . Interesting to know she connects with people from different parts of the world. Such an ambitious woman. I got in touch with Ms. Davis early this year. As a newbie in the market, I had little knowledge on predicting the stock market, but with Ms. Norman weekly analysis and advise profits are guaranteed! I received three times my initial deposit in a week!!
I have a master in mathematical finance, so it wasn’t so easy to get me convinced to begin an investment without me carrying out proper research on her. I had her broker ID checked and she’s fully verified! So I began with a few bucks, only to get huge returns in a month. I reinvested and now I get long term monthly returns… So glad I don’t rely on pay checks anymore.
This is still a window-shopping market. But there are a lot of intriguing stocks to watch from a variety of sectors. You don’t have to act on every forecast, hence i will suggest you get yourself a financial-advisor that can provide you with entry and exit points on the shares/ETF you focus on.kudos to Norman, great remarks!
Great interview ! !
Small constructive feedback: when Geoff Hinton isn't talking the video shows the "Eye on AI" Logo and (for some reason) that's distracting.
I think that this will open so many possibilities.
When working with small MLPs RELU is rearely the best activation function, something like tanh tends to perform much better but if you try to have more than 4 or 5 layers backpropagation chokes on it due to vanishing gradients but with this it wouldn't matter.
It doesn't really have to be an exclusive or between Forward-Forward and back propagation, you could train many small backprop networks and join them with the forward-forward algorithm. It won't be as efficient as forward-forward for an analog hardware implementation but it would likely squeeze more into the same amount of weights and will likely provide better accuracy in some tasks. It will also be much less memory demanding than trying to do backprop over the full network and that would increase what our current hardware can do by a lot.
Backwards connections would be much more trainable even without the trick of replicating the input data and the layers. With true backwards connections, it may still not converge into a stable solution due to the feedback loop formed, but it won't have the issues of backpropagation through time. If that can be made to work, models can develop something akin to our working memory.
Not needing a differentiable model of everything opens the possibilities of inserting stuff in the middle of the network that wouldn't be easy to integrate normally, like database queries based on the output of previous layers or fixed function calculations.
It makes so much sense intuitively that it's hard to comprend that it took so long for this idea to hatch. Hilton is a genius.
Hopefully our AI children wont be this dumb
@@noomade Yes, especially when he tells us AI will kill us all; something he created.
This is a real ai UA-cam channel. I'm sick of all the channels feeding on buzz and popularity over éducative content
Such a good talk, thank you for organizing this Eye on AI! I have been implementing the FF-algorithm in python and whilst the training is understandable, the testing becomes tricky for multi-class classification trained with the supervised version that Hinton describes. This is because for new examples you don't have any labels, so you need to impose all possible labels on top of the example the same way as in the training and run the network with these to see which has highest hidden layer activation or "goodness" as Hinton describes it.
Since the overlayed label is a part of the input, it contributes to the activations, meaning that there is currently no way to test all possible labels at once, which yields to scaling problems for ImageNet or other classification problems with a big amount of possible predictions where every possible class label representation has to be overlayed with the tested input. Will be interesting to see if this can be overcome or if unsupervised learning will be the standard procedure with this technique.
Another super-interesting part in my opinion is the fact that Spiking Neural Networks have the Heaviside function as the activation which has no derivative. So traditionally trained SNN's have a Heaviside forward pass and a Sigmoid backwards pass to tune the weights, using FF we will be able to tune SNN's without having to "trick" the backwards pass to not be a step function, which may yield a better representation of our biological processes.
A.I. and WW3 Updates: REPUGNICANS WANT WW3 & CIVIL WAR IN U.S. - AI WILL GIVE IT TO THEM!! Don’t believe it? Ask AI! (We did!)
“Commercial Artificial Intelligence” implementations (i.e., Enterprise-wide, mature instantiations) will be very bad for global and local economies, easily replacing all workers, including designers, architects, programmers, analysts, writers, accountants, testing and diagnostics, etc, etc, etc. - ALL (expensive) white collar jobs are the soonest at risk.
A.I.-centric CEO’s will MAKE millions being the first to quickly replace all workforce ASAP, starting in the next 12-18 months, as AI “utilities”, then full-blown AI systems and deployments become ubiquitous.
It will occur very quickly in the USA.
*******Even (especially!) CEO’s will be replaced.
Simply put: Using existing corporate data stores and database systems, in the next 12-18 months AI will re-engineer whole economies. Changes will then be implemented, effecting whole market sectors, literally, over night.
Only low level, manual labor skills will be highly coveted but, as the global economy crashes, the result will be scaled down work forces everywhere.
In the USA, it will become very violent, as ignorant people CONTINUE to lose their jobs with no place to turn for work.
Putin’s wartime exit strategy is based on global ollapse to protect his insanity.
Xi will sit back and observe, allowing Kim Jong Un to act as a chess board pawn. Kim Jong Un is an angry psychopath, worse than Putin.
A.I.: THE WEALTHY ELITISTS’ CRACK PIPE
Nearly completed and hoping to keep U. S. distracted, today, the REPUGNICAN’s stinging strategy is more clearly evident, as REPUGNICAN handlers bribe and cajole old and new minions while their elitist controllers are greedily grasping for their newest crack pipe:
*******Native mode Artificial Intelligence used to replace the human white collar and blue collar labor forces, as the early robber barons boldly proclaimed and contemplated, aloud. *******
I DARE YOU TO GOOGLE IT!
Robber Baron, Jay Gould, repugnantly proclaimed as their elitist goal to control the world and rape Mother Earth to extract her finite resources:
**********We will “employ half of America to kill the other half” - Google it.******
We DARE you to seek these (and other) truths!
Another greedy psychopath and Gould contemporary, Cornelius Vanderbilt declared, “What do I care about the law. Ain't I got the power?”
Google it! And then ask your favorite AI chat bot:
Were these well known elitist statements sane or were they the words of a psychopath?
Ask soon! Because elitists control all AI technology and future A.I. implementations which are being hacked, and future versions will soon filter (mask) these early conclusions and edit the truth out of and away from their truth-filled responses, as elitists re-program AI bots to omit truth and, instead, invoke the will and desire of REPUGNICAN strategists!
Basically, it is training a neural network but instead of using positive training data, we're using negative training data. This can yield high perplexity due to the fact no one can get "perfect negative data" but we can easily get positive training data; thus I think it will not replace back propagation, but will be very useful in many applications, like neuromorphic hardware; or maybe even applications where we don't even know what the positive data should look like! So we're reverse-solving the problem somehow. This is really very interesting.
Great interview! I could do without the blinking eye thing.
Fantastic interview. I may well need to listen to it 3 or 4 times!
I wonder how the forward algorithm, capsules and "GLOM" connect to building those "world models" from observation. I think I understand Yann when he says that you shouldn't make generative models that predict things like pixels, but make predictions about more abstract representations so that you can ignore irrelevant details (like leaves blowing in the trees). Making predictions about higher order, more abstract concepts like "which car overtakes who" etc will make the network start modelling dynamics, and gain an understanding of what it sees, including causal reasoning. Is this Hinton's plan too or does he not think in terms of world models?
this is obvious, real question is how to decide what's relevant and what's not, then this will change with time when system learns new concepts and so generative models have to change, how to make such system stable?
Schumachers Batman, see the Yann interview I just posted. He addresses your question obliquely.
Extremely fascinating to hear this after Chomsky's criticisms of the current deep learning paradigm as failing to differentiate between possible and impossible languages
Chomsky is a dumbass
@@phoneticalballsack why do you say that?
@@AZTECMAN Have you talked to him in person?
@@phoneticalballsack Nope. But my lack of personal encounter doesn't seem very important to understanding your statement.
Please explain to me, why Chomsky is a dumbass. If you happen to have met him, I'd certainly welcome a anecdote though I don't consider it crucial.
The discussion at 33:32 immediately suggests the possibility of applying a "color" to each neuron, where the squared activation of neurons of one color contributes positively to "goodness", and the squared activation of neurons of the other color contribute negatively to goodness. Any given layer could have neurons of *both* colors.
Of course, that leads to additional questions:
1. Is there a rule for determining each neuron's color that could be applied a priori to give better results?
2. Should there be a rule for changing/updating the color of a neuron so the distribution of colors can be adapted to the problem and the data at hand?
Finally, to get even farther afield: something whose activation squared counts as positive sounds like a real number. Something whose activation squared counts as negative sounds like an imaginary number. Instead of choosing between one of two colors for neurons, should the activations be multiplied by a *complex* number before squaring, with the sums of the real parts of the squares being used for the objective? Because the effect of complex color is continuous and differentiable, it may be trainable. The network could find, through learning, the balance of importance between features and constraints for the problem domain.
Fascinating discussion! Thanks so much for posting it, and extra thanks to Prof. Hinton! He explains things very clearly.
Could be a historical interview for all time in the future. Good job.
I think the idea of high layer-activations only for the positive data, interesting. The network essentially isn’t giving an Output like in backpropagation, but it’s now the Property of the network to “light up” for correct labels, and therefore indicating whether it’s a positive data or not. I enjoyed this interview given by Hinton about his paper.
Great to we can hear Dr. Hinton's lecture through social media.
25:18 hidden layer is asking: "are my inputs agreeing with each other, in which case I'll be highly active, or are they disagreeing, in which case I won't." :)
Thank you for this interview. Though I don't understand the technical details of it, I did get to draw on some simple things, and also was able to appreciate the serious brain power in Mr. Hinton.
There are other ways to achieve what backprop does, without backprop: use complex, not linear, quantities; use Conversation Theory; use Active Inference.
"Attenuation" is a term used by neurosciences for enforcing the "fake data" / "real data" discernment.
Would negative data training be somewhat similar to hypothesis testing? Or at least what they originally conceptualized a null hypothesis as but has now been obscured. Trying to maximize true negatives as opposed to minimizing false positives.
이렇게 재테크 유튜브중에 가장 가슴에와닿고 고갤끄덕이게하는영상이 있다니!!!
Fascinating model. His view of consciousness doesn't seem as good as Joshua Bach's work though. He says there are a million definitions of consciousness but I believe the most commonly used meaning by philosophers says consciousness is the feeling that its like to be something. Consciousness is a model of a person embedded in a story generated by the neocortex to be stored in memory.
see what Yann says about consciousness in the latest episode: my full theory of consciousness ... is the idea that we have essentially a single world model in our head. Somewhere in our prefrontal cortex and that world model is configurable to the situation we're facing at the moment. And so we are configuring our brain, including our world model for ... satisfying the objective that we currently set for ourselves.
... And so if you have only one world model that needs to be configured for the situation at hand, you need some sort of meta module that configures it, figures out like what situation am I in? What sub goals should I set myself and how should I configure the rest of my brain to solve that problem? And that module would have to be able to observe the state and capabilities - would have to have a model of the rest of itself, of the agent, and that perhaps is something that gives us the illusion of consciousness.
@@eyeonai3425 it's models all the way down!
What was the constraining (low variance) complement to PCA Hinton mentioned?
Makes me wonder. Do things like LSD perhaps trigger parts of this 'sleep' state system, but while still awake. Makes quite a bit of sense to me, especially considering how extremely similar 'tripping' hallucinations are to the things AI produces when it is allowed to 'dream away'. Curious.
im high af watching this video and im like, "hooooly shit this vid is one trippy dope" lol
Oh man the eye was intense! Your all relaxed listening to Hinton's genius and then BAM! a giant spooky eye appears out of nowhere and scares the bejesus outa ya.
Intense experience.
I have been trying to find the podcast where Hinton basically says that the longer length of tokens contributes to hallucinations and variance based on standard ML/DL, anybody out there that heard the same thing?
44:56. I think it depends on what is meant by "but it doesn't really matter if you can't tell the difference." Do we simply mean, as long as the illusion is convincing? Like a Hollywood special effect? Or do we mean, it's not "possible" to tell the difference, because it's beyond our capacity to interrogate? The former is a matter of laziness, where we are willing to accept the "optical illusion" because we don't want to understand the magic. Whereas the latter, the situation has moved to a point where we've pushed the investigation to a sort of "event horizon" from which we are bounded from making any further inquiry. I think it very much matters which of these situations we find ourselves in; ethically, if nothing else.
31:34 capsules, depth in pixels, and comparison to how babies learn, concentrating on what's odd
Such an inspiring interview! But that blinking eye makes me a little dizzy, perhaps I prefer it to be 'static' haha
Hi Jeff. As infant animal learners, we output a behavior and get almost immediate feedback from a parent on whether that behavioral output of a moment ago was "good" or "bad." Did mom look away or smile and interact more? This seems like a crude but fair example of back propagation. No? What do you think Mr. Hinton?
Let me see if I understand
He is redesigning the black box.
Classical black box has explanatory features in the entry and labels or variables to be predicted in the output.
In this approach, everything is in the input. And the output is the "hint of simultaneity" of blocks of entries.
If that's like so, I would like to stress that this concept is the foundation of all this. The learning algo depends on this structure.
One more thought. "Idea association" works this way. "Perception-action" must work in another way. Action looks like an output. Or can it match a FF framework
Really enjoyed the talk but I do wish you’d ditch the big blinking eye. It’s distracting.
Yes it's disturbing please don't do that
@@Xavier-es4gi thanks for the feedback. wont' use it again.
I’m wondering if the brain isn’t using both the positive and negative training at the same time. Much of daily brain operation is on the negative training. Surprise generates activity. Otherwise not active.
Great talk! Also looking forward to see the Matlab code.
I'm also slow at reading especially when it comes to equations!
Multitasking is beneficial for the brain, it mixes things up.
very interesting but not so easy to understand for laymen/women, perhaps another FF Algo video would be very enlightening, thanks God bless.
마음가짐이 정말 중요하죠.
The largest neural network has a trillion connections, which is about a cubic centimeter of the human cortex, which is about 1,000x larger...
What a magnificent thing the human brain is!
But transistors are more than 1000x faster than synapses, in some cases billions of times faster. And smaller.
@@lucamatteobarbieri2493 and for the same amount of computation as the human brain does, uses many times as much energy.
Not a problem for a stationary computer, it'd never work for biological beings even if they were born fully formed for their brains and their sizes.
AI is about to change your world, so pay attention. Love it :)
That blinking eye is really annoying. I'd rather see the interviewer.
This comment is for future visitors! ♥️
I was here! 26 January 2023.
I agree about consciousness. It's a matter of degree, I think and that's what I hear Hinton saying.
12:16 what exactly Hinton means by "negative data"
13:01 supervised learning with an image with correct/incorrect data
14:10 subtracting negative (incorrect) data from positive (correct) data
16:34 example of negative data in a negative phase you use characters that have been predicted already.. you're trying to get low activity cuz it's negative data..
17:04 they cancel each other out if your predictions were perfect (positive and negative phase)
33:11 the very basic algorithm of how to generate negative data effectively from the model should be done nicely before you choose to scale it up
Groundbreaking!
AI/ML only want one thing, and it's disgusting - Hinton's MATLAB code.
Mate don’t generalise your own opinion to the whole AI/ML.
@@artlenski8115 😆
Hinton's Matlab code is disgusting?
@@mmvblog 57:43
Lmao, it’s a meme. Calm down nerds
54:08 Yann LeCun's convolutional neural networks are fine for little things like handwritten digits but they'll never work for real images says the vision community
56:17
When there finally was a big enough data set to show that neural networks would really work well, Yann wanted to take a bunch of different students to make a serious attempt to do the image convolutional neural network work, but he couldn't find a student who'd be interested in doing that :( and at the same time Ilya Sutskever and Alex Krizhevsky, who's a superb programmer, started to be interested in doing that and put a lot of hard work into making it work eventually.. so Yann LeCun deserves to be mentioned, too, according to Geoffrey Hinton
The problem with forward propagation is that it may change its mind to a projection already made and switch fast back again to earlier prediction. However, it is still the better than back propagation. Actually "funny", because negative data is how you get rid of all the BS you don't want to know 🙂
Can someone explain what he means by real data vs fake data? ~7:30 ish
I think he means T and F prediction
I'll never look at a pink elephant quite the same way again 🙂
Zero explanation what "high" vs. "low" activity mean.
it means literally that, high or low magnitude of output vector
Entirely (as whole) the world data is composed by; good, bad, and hallucinating (half good+half bad) data. You can't make non hallucinating AI with current data. Probable far far in future AI will can solve somehow to be non hallucinating. Or you can make one special AI to filter out the hallucinating data, but is not good idea, lot things to work need hallucinating data. My noob opinion.
I have been watching and following this man since 2007 and all I have to say is he is an "EXTREMELY SMART FOOLISH MAN".
Why foolish?
Don’t appreciate your eye flickering eye motif repetition gaining frequency in a disturbing way; instead of staying on the guest… you are open to subliminally training … regardless of intent, IS illegal as well as against u tube regulations. Not good either way. Ive documented. Desist.
A logo talking is so creepy.
Bro just put your logo in the corner or something, no need to flash the whole screen, its just distracting to the conversation
the eye thing popping up is ANNOYING, just stop it
It was annoying to watch so l would have to listen to your program.
It's better to see the person who asks questions instead of some White screen
Geoff chose the wrong acronym. Pink elephant. The N Vietnamese had pink elephants. They rolled in the red clay and became pink. Geoff seems to be taking of absurdity rather than reality. To me pink elephants really are a thing in reality.
8:08
8:12
8:43 similar to GANs
8:51 using the same units
insanely annoying and pointless
Talking without slides is waste of time
That eye is used in superstitions.
Sooooo annoying with that eye 😝
Very enlightening video, There’s this woman I got in touch with during the 2020 lockdown which cost me my job. Ms. Norman Davis helped me manage my assets by introducing my to the best trading platform and strategies, I earned a lot of $$$ working with Norman at the comfort of my home. I still keep in touch with the amazing lady
Hi there, I’m commenting from Switzerland . Interesting to know she connects with people from different parts of the world. Such an ambitious woman. I got in touch with Ms. Davis early this year. As a newbie in the market, I had little knowledge on predicting the stock market, but with Ms. Norman weekly analysis and advise profits are guaranteed! I received three times my initial deposit in a week!!
I have a master in mathematical finance, so it wasn’t so easy to get me convinced to begin an investment without me carrying out proper research on her. I had her broker ID checked and she’s fully verified! So I began with a few bucks, only to get huge returns in a month. I reinvested and now I get long term monthly returns… So glad I don’t rely on pay checks anymore.
I’m trying to create long term wealth to set towards property one day . How can i reach out to her? she could be of great help
< normandavis
This is still a window-shopping market. But there are a lot of intriguing stocks to watch from a variety of sectors. You don’t have to act on every forecast, hence i will suggest you get yourself a financial-advisor that can provide you with entry and exit points on the shares/ETF you focus on.kudos to Norman, great remarks!