"Oh that's what I've been thinking for almost 20 years." I've been following Bengio's work for awhile now and this was truly an incredible conversation, in terms of both answers by someone patiently tackling fundamental problems in the field and questions by a group that has been patiently studying the field and connecting the ideas. Mad props!
This was really satisfying in terms of both questions and answers. So I can't even think of any question (or answer) not already given. We have our work cut out for all of us. Now you must have Jeff Hinton to complete the "Deep Learning Trio".
This one was truly amazing! Imagine being Bengio, studying these topics for decades, replying to every question in hundreds of talks around the world, working with groups of bright people, and one day having a conversation for a youtube channel where he clearly enjoys and smiles at every question, to the point where he needs to say it out loud. It was like seeing Goku enjoying an amazing fight. The good vibes on this one were flowing both ways (pun intended), not only uncovering a better understanding of the topic but also motivating the viewer to learn more about it. Congratulations guys! This talk was pure delight.
Another great show. I liked the intro a lot, that put me in the right frame of the mind for the following discussions. Prof. Bengio has scientific ways of thinking and explains various problems in AI and machine learning. I learned a lot.
I definitely need to read the papers about what they are calling "causality" here. It sounds very promising. Thanks for taking the time to post links in the description.
Hell yeah! You have a merch store and patreon! Sweet that I don’t need to feel guilty about enjoying these insane quality videos and only supporting you with likes and comments hahaha
Thanks for great conversation. Hope someday understand it fully and make it actually work >D I hear same things over and over again. Feels good to know direction of the field in some sense.
You guys deserve the props from bengio, well done! I would love to hear more discussion and dissection of this “free lunch if there’s structure” notion. How exactly are we beating dimensionality and combinatorial vastness? The answer has to be abstraction arising from, or baked into, the architecture + algorithm. But if this is truly effective at scale, then it implies weighting for exploration converges on parsimonious modeling- so that you can get powerful and versatile composition of abstractions. This is interesting: it associates “casting about” with “finding the best explanation” (which presumably tends to generalize or transfer well). Sort of turning exploitation on its head, no? (The conversation kept circling around the attractors of GOFI theme of composing abstractions, and information as a reward.) Honestly I deeply appreciate the irl oracle question search aspect, but I think Keith is on to something with the flowZero line of thought. It would be informative to understand learning rate and policy space in a setting where we already have some kind of grasp.
How can you have abstractions without understanding first? A NLP system doesn't know what the text mean as the text is referring to something inaccessible to it(external world and its entities and dynamics), Can you recover the full information from just text? Obviously not.
Actually, you can! But that depends on, well - semantics, that is how you define what abstraction is. For example, so called abstract art does not (always) builds on understanding. Rother it interacts with understanding, some times changes it and vise versa. It is more constructive to speak about generalization and systematization since both have semantics at least partially defined. This why I always respond to people saying that LLM like GPT-3 do generalization - not to confuse generalization with degeneration - that is simplification.
Current models already aren't forced to get information from text alone, but a recent paper titled "One model for the learning of language" shows you can indeed learn semantics from observing a language and do that quite fast.
A simpler idea for making longer predictions from videos and chronology with what already exists ! - by Marilyse Devoyault, 1. Make time lapses from big data of available videos. For each image obtained, link : shooting number & real time date-hour of image : Ex: from shooting #789 2022-04-22 20:38:01 2. Use classifier to identify main elements of the identified context at the beginning (some pictures at the beginning with detailed elements to tell what to look for) Let say you have a DallE2 give you a first image of a dog, a cat and a tree. You take this picture to input in your new predictor with this classifier. It could also be a robot taking a picture of a new situation in front of him. 3. Have this predictor find every picture in the data that is close to your picture 4. Have some type of transformer find the next picture (using number of shooting date-hour) of every picture you found in 3. 5. Use some king of GFlow net to regroup similar next pictures found in 4 and keep the main probabilities. 6. For example, two possibilities : step 3, a picture with a dog seeing a cat and a tree nearby. Step 4, many next images possibles : cat runs toward tree and dog chases; cat turns toward dog and spit; cat lay down and dog happy, dog go away and cat don’t move; bird comes from the tree and land on the dog head… Step 5, the GFlownet regroup all the images of the cat running toward the tree, and all the images of the cat facing dog and spitting. Since they are numerous, they are the main probabilities kept for the predictions and to go on with next prediction 7. Take the images from the main flows of 5 and find the next picture (using number of shooting date-hour) of every picture you found in 5. 8. Use the GFlow net to regroup similar next pictures found in 7 and keep the main probabilities. 9. Go on as long as you can to make a longer plausible prediction with main probabilities of what can happen.
Tim, can you square how you seem to react positively to YB's proposal at the one hour mark that this mini world model could give us the "illusion of Cartesian Dualism" but then are very negative on Ilya's "a little bit conscious" comment on his Transformer enable architecture. They sound like similar ideas, no?
lol @ the twitter reference. As somewhat of a Twitter addict myself recently, imagine... to just *imagine* the productivity improvement from shunning Twitter. It is unfathomable.
Fixing a set of dimension reductions? Eventually expanding a split into a small dimension which would split any min/max. So fix y, solve, set y=0.0000x ... and flow from a 1D to a 2D ...?
Do any language models take their own output and feed it back in with the next input from the human? The thought here being that currently these bots are having a weird sort of interaction where they are only fully aware of one half of the conversation
The financial market has tough one this past months, but I watch interview on CNBC where the anchor kept mentioning "...CATELYN MORRIS...". This prompted me to touch with her, and from October 2021 till now we have been working together, and I boast now of €35k in my trading wallet.
this is the miss conception going around, you call it love for money while some see it as receiving good information, which can still be a miracle in the making by God, how has the father worked miracles since the days of abraham it has always been through men
@@harriswealthers1579 lots of people forget this, they are still waiting on manner to fall out of the skies personally I pity such people they have been brainwashed by society that things are meant to given to them which is wrong you find things and work for them either by getting cautious with sensitive info and knowing which info to act on
I don't think neuroscience will ever get an explanation for how consciousness arises. Matter, this abstract, contour that can be fully described by quantities (completely devoid of qualities) and subjective experience are two incommensurable categories. The best you will get is "and poof, consciousness arises".
Oh no! Mr Stanley stired my artistic fiber… A question to GFlowNets
The head as a materialised Plasma ball Every impulse from the outside Creates a pattern in the grey matter, chronologically impregnated. Monday 7 o’clock in the morning, I see my pineapple on the counter. This creates a pattern looking like a Dryopteris in my visual cortex. Monday 7 o’clock at night, I taste some new basswood herbal tea. This creates a pattern looking like an Osmunda in my gustatory cortex. Tuesday 7 o’clock in the morning, I see my pineapple on the counter. It isn’t ripe. This creates another pattern looking just like a Dryopteris with a slightly different shape in my visual cortex, This pattern is infinitely close to the Monday 7 o’clock am pattern of a Dryopteris, but absolutely not the same. It is impregnated almost at the same place in my grey matter, but infinitely slightly more inside. The discrepancy is infinitely small. Tuesday night the taste of my basswood herbal tea and its Osmunda pattern will slip right next to my Monday night slightly different Osmunda pattern. More inside. I am grasping the taste of basswood herbal tea. Could it be how everything is chronologically impregnated by electric impulse? Trillion of trillion of microscopic layers? Could it be how we can make predictions? Since everything is chronological in my materialised plasma ball? Could it be how we generalise with infinitely small layers (chronological layers) of tiny hair of patterns that are alike and almost merge, straw inside a straw inside a straw inside a straw, but when this grey matter area is visited by an impulse, flows the general concept of a pineapple or basswood herbal tea? GFlowNet, will you learn to consider time? Will you make chronological layers of flows? Will you learn the chronology of your encounters so that you may imagine the future? Will you use the flows with numerous layers to grasp the Platon Idea and it’s relation to time, in other words, its probabilities to exist following the happening of a previous Platon idea? Between two dense areas of layers of straws, are most of the straws of one area pretty much from the same impregnated impulse jet of the other area of dense layers of straws? Yes ? Then we have a dialectic!
He didn't really have an answer for: [01:01:25] Why are linear models dominating? They are abstraction! However there are answers out there, even if they require a certain human psychological reboot to move drastically to a new vantage point on the problem. Maybe poke around on archive dot org with ReLU as a switch.
The same way when we approximate derivative as the linear slope of two dots when they are very close to each other, also similarly in calculating integral. We can approximate a curve with a bunch of linear lines. Relu creates a kick of non-linearity on top of a linear function, makes it even easier for neural nets to carry out the task. It's much easier to learn linear functions than less defined non-linear functions.
Oh yes! This one is a BANGER folks -- ENJOY!!
"Oh that's what I've been thinking for almost 20 years."
I've been following Bengio's work for awhile now and this was truly an incredible conversation, in terms of both answers by someone patiently tackling fundamental problems in the field and questions by a group that has been patiently studying the field and connecting the ideas. Mad props!
Can't even describe how excited I am about this one! MLST strikes again, thanks guys!
the explanation in the begging was great, and prof. Bengio made it more discernable.
This was really satisfying in terms of both questions and answers. So I can't even think of any question (or answer) not already given. We have our work cut out for all of us. Now you must have Jeff Hinton to complete the "Deep Learning Trio".
Geoff
What a wonderful interview. Prof. Bengio is honest and brave with class.
You guys and yannic Kilchers channel are really such good sources to get to know awesome research topics. You guys are the best!
This one was truly amazing! Imagine being Bengio, studying these topics for decades, replying to every question in hundreds of talks around the world, working with groups of bright people, and one day having a conversation for a youtube channel where he clearly enjoys and smiles at every question, to the point where he needs to say it out loud.
It was like seeing Goku enjoying an amazing fight.
The good vibes on this one were flowing both ways (pun intended), not only uncovering a better understanding of the topic but also motivating the viewer to learn more about it.
Congratulations guys! This talk was pure delight.
Thanks Jack, we really appreciate it!
Thank you so much! We are very proud of this one and so very fortunate to have had him on the show.
Another great show. I liked the intro a lot, that put me in the right frame of the mind for the following discussions. Prof. Bengio has scientific ways of thinking and explains various problems in AI and machine learning. I learned a lot.
I definitely need to read the papers about what they are calling "causality" here. It sounds very promising. Thanks for taking the time to post links in the description.
OMG, is this real? 🤯 Awesome, can't wait to watch it whole!
Great to hear you have loads planned for this year
Amazing video introduction very well done and I really dig the music it put me into a meditative state
Wooow, the production value is impressive! Listening to the audio version is not the same
Hell yeah! You have a merch store and patreon! Sweet that I don’t need to feel guilty about enjoying these insane quality videos and only supporting you with likes and comments hahaha
Thanks for great conversation. Hope someday understand it fully and make it actually work >D I hear same things over and over again. Feels good to know direction of the field in some sense.
MLST is the best!!
Thanks! We really appreciate our amazing audience 😎
On a roll!
It might be interesting to invite Judea Pearl and ask his opinion about the possibilities for AI to grasp causal inference.
if this is not one of the best episodes, I don't know what is.
You guys deserve the props from bengio, well done!
I would love to hear more discussion and dissection of this “free lunch if there’s structure” notion. How exactly are we beating dimensionality and combinatorial vastness? The answer has to be abstraction arising from, or baked into, the architecture + algorithm. But if this is truly effective at scale, then it implies weighting for exploration converges on parsimonious modeling- so that you can get powerful and versatile composition of abstractions. This is interesting: it associates “casting about” with “finding the best explanation” (which presumably tends to generalize or transfer well). Sort of turning exploitation on its head, no? (The conversation kept circling around the attractors of GOFI theme of composing abstractions, and information as a reward.)
Honestly I deeply appreciate the irl oracle question search aspect, but I think Keith is on to something with the flowZero line of thought. It would be informative to understand learning rate and policy space in a setting where we already have some kind of grasp.
Mindblowing!
Wow! Epic video! 😍
the best podcast out there... Keep it up!
Well done. I'm looking forward to your interview with David Chalmers and his reaction on the 'awareness and experience' of an AI algorithm.
How can you have abstractions without understanding first? A NLP system doesn't know what the text mean as the text is referring to something inaccessible to it(external world and its entities and dynamics), Can you recover the full information from just text? Obviously not.
Actually, you can! But that depends on, well - semantics, that is how you define what abstraction is. For example, so called abstract art does not (always) builds on understanding. Rother it interacts with understanding, some times changes it and vise versa. It is more constructive to speak about generalization and systematization since both have semantics at least partially defined. This why I always respond to people saying that LLM like GPT-3 do generalization - not to confuse generalization with degeneration - that is simplification.
Current models already aren't forced to get information from text alone, but a recent paper titled "One model for the learning of language" shows you can indeed learn semantics from observing a language and do that quite fast.
@@lenyabloko You can do it a degree, can you understand smell from text on the web?
That title though. Bet it would be a great talk!
joshua benshi is a rock star of ml
Awesome work guys 😋. Please interview Max Tegmark!
Awesome episode :)
A simpler idea for making longer predictions from videos and chronology with what already exists ! - by Marilyse Devoyault,
1. Make time lapses from big data of available videos. For each image obtained, link : shooting number & real time date-hour of image :
Ex: from shooting #789 2022-04-22 20:38:01
2. Use classifier to identify main elements of the identified context at the beginning (some pictures at the beginning with detailed elements to tell what to look for) Let say you have a DallE2 give you a first image of a dog, a cat and a tree. You take this picture to input in your new predictor with this classifier. It could also be a robot taking a picture of a new situation in front of him.
3. Have this predictor find every picture in the data that is close to your picture
4. Have some type of transformer find the next picture (using number of shooting date-hour) of every picture you found in 3.
5. Use some king of GFlow net to regroup similar next pictures found in 4 and keep the main probabilities.
6. For example, two possibilities : step 3, a picture with a dog seeing a cat and a tree nearby. Step 4, many next images possibles : cat runs toward tree and dog chases; cat turns toward dog and spit; cat lay down and dog happy, dog go away and cat don’t move; bird comes from the tree and land on the dog head… Step 5, the GFlownet regroup all the images of the cat running toward the tree, and all the images of the cat facing dog and spitting. Since they are numerous, they are the main probabilities kept for the predictions and to go on with next prediction
7. Take the images from the main flows of 5 and find the next picture (using number of shooting date-hour) of every picture you found in 5.
8. Use the GFlow net to regroup similar next pictures found in 7 and keep the main probabilities.
9. Go on as long as you can to make a longer plausible prediction with main probabilities of what can happen.
1:09:11, I have the same thoughts on the matter
Tim, can you square how you seem to react positively to YB's proposal at the one hour mark that this mini world model could give us the "illusion of Cartesian Dualism" but then are very negative on Ilya's "a little bit conscious" comment on his Transformer enable architecture. They sound like similar ideas, no?
"we had a paper, i think it was at Neurips" ... lol he's so successful he can't even remember what he's published at neurips
🤣
lol @ the twitter reference. As somewhat of a Twitter addict myself recently, imagine... to just *imagine* the productivity improvement from shunning Twitter. It is unfathomable.
Quicker than a what now?
"Whippet with a bum-full of dynamite" ;)
On the topic of consciousness, I just finished The Case Against Reality by Donald Hoffman. It would be great if you could invite him to the MLST show.
Great idea! We would love to have him on
@@MachineLearningStreetTalk also bernardo kastrup
Thats me :D:D:D:D.
MLST: Bengio, you're the best!
Bengio: No U!
What a lovefest
1:16:13 well than we‘re both suffering 😂 oh boy
Sounds like a high gradient is well changy.
Estimated maximal possible differential between d2/dx2 maxima and minima, for placement of secandary "randomization" betterment?
Fixing a set of dimension reductions? Eventually expanding a split into a small dimension which would split any min/max. So fix y, solve, set y=0.0000x ... and flow from a 1D to a 2D ...?
A hyper cube in 1D is just a line.
Add the last latent layer 1 everything neuron to many delta neurons one by one at a time?
Is delta a latent causality? Changing one dimension changes all the latent delta nuron outputs?
Wow
Who is the guy at the bottom right?
Keith Duggar
Uh oh ... it's me. Dare I ask why?
Do any language models take their own output and feed it back in with the next input from the human?
The thought here being that currently these bots are having a weird sort of interaction where they are only fully aware of one half of the conversation
Maybe it would require something like a GAN to pull it off, and also maybe if it did the network would have some notion if self
Hell yeah!
Discord link?
discord.gg/HNnAwSduud
wondeful
The financial market has tough one this past months, but I watch interview on CNBC where the anchor kept mentioning "...CATELYN MORRIS...". This prompted me to touch with her, and from October 2021 till now we have been working together, and I boast now of €35k in my trading wallet.
this is the miss conception going around, you call it love for money while some see it as receiving good information, which can still be a miracle in the making by God, how has the father worked miracles since the days of abraham it has always been through men
Your English may be poor even if your intuition to deduce money methods are impeccable.. .how were you able to get a meet .this is rare
@@harriswealthers1579 lots of people forget this, they are still waiting on manner to fall out of the skies personally I pity such people they have been brainwashed by society that things are meant to given to them which is wrong you find things and work for them either by getting cautious with sensitive info and knowing which info to act on
@@lucyweilbel6681 with norristrades as the user name .. we talk on the t e l e g ram better to ask madam you'rself
@@sarah-lp2oc your english is funny 😅😅
Diversity, baby!
I don't think neuroscience will ever get an explanation for how consciousness arises. Matter, this abstract, contour that can be fully described by quantities (completely devoid of qualities) and subjective experience are two incommensurable categories. The best you will get is "and poof, consciousness arises".
Feel sad for the guy on Yannic's discord who watched 1/3rd the way through and stopped. At just about that point it goes from 0 to 100 real quick.
pog
Oh no! Mr Stanley stired my artistic fiber… A question to GFlowNets
The head as a materialised Plasma ball
Every impulse from the outside
Creates a pattern in the grey matter, chronologically impregnated.
Monday 7 o’clock in the morning, I see my pineapple on the counter.
This creates a pattern looking like a Dryopteris in my visual cortex.
Monday 7 o’clock at night, I taste some new basswood herbal tea.
This creates a pattern looking like an Osmunda in my gustatory cortex.
Tuesday 7 o’clock in the morning, I see my pineapple on the counter. It isn’t ripe.
This creates another pattern looking just like a Dryopteris with a slightly different shape in my visual cortex,
This pattern is infinitely close to the Monday 7 o’clock am pattern of a Dryopteris,
but absolutely not the same.
It is impregnated almost at the same place in my grey matter, but infinitely slightly more inside.
The discrepancy is infinitely small.
Tuesday night the taste of my basswood herbal tea and its Osmunda pattern will slip right next to my Monday night slightly different Osmunda pattern. More inside. I am grasping the taste of basswood herbal tea.
Could it be how everything is chronologically impregnated by electric impulse?
Trillion of trillion of microscopic layers? Could it be how we can make predictions? Since everything is chronological in my materialised plasma ball?
Could it be how we generalise with infinitely small layers (chronological layers) of tiny hair of patterns that are alike and almost merge, straw inside a straw inside a straw inside a straw, but when this grey matter area is visited by an impulse, flows the general concept of a pineapple or basswood herbal tea?
GFlowNet, will you learn to consider time? Will you make chronological layers of flows?
Will you learn the chronology of your encounters so that you may imagine the future?
Will you use the flows with numerous layers to grasp the Platon Idea and it’s relation to time, in other words, its probabilities to exist following the happening of a previous Platon idea?
Between two dense areas of layers of straws, are most of the straws of one area pretty much from the same impregnated impulse jet of the other area of dense layers of straws? Yes ? Then we have a dialectic!
He didn't really have an answer for:
[01:01:25] Why are linear models dominating? They are abstraction!
However there are answers out there, even if they require a certain human psychological reboot to move drastically to a new vantage point on the problem.
Maybe poke around on archive dot org with ReLU as a switch.
The same way when we approximate derivative as the linear slope of two dots when they are very close to each other, also similarly in calculating integral. We can approximate a curve with a bunch of linear lines. Relu creates a kick of non-linearity on top of a linear function, makes it even easier for neural nets to carry out the task. It's much easier to learn linear functions than less defined non-linear functions.