วีดีโอ

thank you for much for this amazing lecture

What is the meaning of no closed form in step 2 of Bayesian inference?

Thank you

Wonderful explanation about ABC!

Somehow glucose is more efficient than electricity. 😂

Nice conference

I think i have bettered the understanding on Yashua Bengio on consciousness and meta learning and why deep learning depedendce on bits for computing and inference search algorithm of gradient descent are its limitations to what it can do.

Yesterday i had given a comment on twitter on contrastive learning, i would like to correct my statement that there is dimensionlity redcution but it is in the form of a concept. So, basically its just another way of looking at transfer from an entropy perspective. Comparable entropic systems are similar, hence, i don't see many conceptual different between what transformer q, k, v do and energy function do. Its just another good name.

Hinton' view of vision as hierarchical is not true. In perception, hierarchy does play a role but hardly in vision. Vision is what a 3D mapping by eyes of an external environment which is used as data for infinite dimensional decision making. World exists in infinite dimesnional space of physical planes. Eyes are evolved to create a 3d physical mapping of these planes and passed to human perception engine which makes sense of this using physical and meta physical spaces where human perception wander, although in a very optimized manner(much much better than computing) to find optimized paths or decision making. Those who excel in finding paths which are best are often people excelling in those dimensions whether its chess, football, occult, porn, movies or whatever. However, enhancing a dimesnion is like writing the best function for this nuance. We live in a world of recursion as well as non-recursion of several functions which are either a composition, association or hierarchical. All at the same time.

The biggest problem in computer vision is not specific to machine learning and neural networks, as opposed to the inherent way images are encoded for computer processing. This starts with the way most camera sensors work to convert light waves into pixels using the bayer filter. And the output of that is simply a file with a bunch of pixels which are nothing more than 3 numeric values representing red, green and blue color values for each pixel. These numbers are not light waves that are converted by the eye into bio chemical electrical changes as they move from the retina into the visual cortex. And when humans see images on a computer they don't see numbers they see the light waves projected by the monitor or display device. All of that is a big problem in computer vision in many ways, from color perception, to pixel processing, segmentation and so on. For colors the primary problem is that when your eye looks at an image it is very easy to pick out specific color regions, such as a red apple in a bowl of fruit, based on light spectrum values received in the retina. A computer cannot do this because each pixel has a unique r,g,b value and there is no simple way to define what groups of pixels represent a specific color and every pixel has "red" in it. Which means that for computer vision most of the processing involves coming up with some way of determining the relationship between each pixel and the pixel next to it. This is generally called a convolution where you go through a region of pixels and apply a mathematical operation in order to "simplify" it for purposes of generating a statistical relationship over a group of pixels. But that still isn't the same as seeing a "red" apple in the bowl of fruit like the human eye. And that is because the human eye can easily group various light signals together as part of coherent shapes and objects. This grouping is called segmentation in computers and is a difficult task that depends on using some kind of convolution algorithm to try and assign mathematical relationships for this kind of grouping based on large volumes of data. And this is where CNNs come into play for computer vision, but these CNNs do not provide the same ability to process visual signals as humans do because the human visual system has no concept of pixels or convolutions.

Brilliant. Great teacher.

It's should rationally be obvious that to understand artificial neural nets you should understand the dot product. However apparently the weighted sum is a summie thingie. And this is the state of neural network research in 2021.🎃 AI462 neural networks.

Need a better mic...

10:31 what is it you do with the coordinate frame??

"Stacked Capsule Autoencoders" paper has got 37 citations so far. I think Geoff might be onto something here hahaha (not being sarcastic!). Nice presentation! I especially liked the Q/A part. Their presentations were way too abstract for a broad audience as you'd have to have some more context about their particular research area to get something super valuable out of it. Coming from a guy who does this for living in Microsoft and also in his free time. So pretty much all the time I'm awake. We need more fireside chats with AI minds on TH-cam. As for the keynotes I got from this talk: attention and self-supervision will play a significant role in the future of ML research. Also, "crazy" ideas like capsule nets or more generally questioning the current methods is super important - we need researchers who are working without overfitting to h-index as their main metric and instead focus on the long-term goals of improving ML for the greater good.

Thank you for posting this. I loved the Q&A session and even took some notes, it's very intimate and inspiring!!

Can someone use machine learning to clear the damn audio

I like his way of explaining things....truly intuitive

Africa is a country

The introduction starts @3:47 Speaker starts @4:41

12:05 Control theory need not concern itself only with continuous dyn. systems. When you talk about continuous vs. discrete time, for example you're just changing the 'time' set you work with, i.e. a countable infinite set (set of integers, lattices, etc.), and an uncountable dense set (the reals). When you talk about the state-space being continuous or discrete, you talk about whether its a uncountable or countable set, respectively, and even finite countable (state machines, Markov chains, etc). All of these concepts are covered in standard graduate textbooks about dynamical systems and control theory. To claim that control theory does not cover discrete, and RL does is just false.

To Yann Le Cunn: learning to plan complex action sequences actually are very easy to solve. I have a solution

Bodyguard: How tough are you? Geoffrey Hinton: I classified MNIST digits with 98.7% accuracy. Bodyguard: Yeah? That's not tough. Geoffrey Hinton: Without using the labels. Bodyguard: Sir, here you go, sir. Sorry, sir.

Did they decide that all of them will hit the mic?

I found the better HD resolution of the same video:- facebook.com/icml.imls/videos/432574700590384/

Are the slides available?

what a historical moment. and I was there. what a magic moment to be alive in history of AI

I had to download the video and watch it in VLC, 'cause I couldn't hear a thing

1:56:39 Le Cunn dodged that like Neo, but it came back haha

Hinton's explanation of transformer is the best I've seen so far

What a time to be alive. This is inspiring :) I really wish they had answered the question at the end about The Measure of Intelligence paper. We need better ways to measure AGI and better datasets to train it on.

Pure joy to watch

have the video of them receiving the Turing award come out yet?

Thank you for sharing!!!

CNNs are rubbish! lol

Capsule for 2020. Btw for first comment.

According to the definition of this guy of the local optimum, x = 0 is a local optimum of the function y=x^3 !!! 15:15

Slides and Bibliography unsupervised.cs.princeton.edu/deeplearningtutorial.html

26:03 SBEED Convergent RL with Nonlinear Function Approximation.

Good explanation, but why is this method called "variational" Bayes ?