Early view of next Neural Networks video

Great video! How do you animate the 3d graphs?

Great video! One grammatical nitpick: at 12:31, you use the phrase, "This is why, by the way, that artificial neurons have ..." If you took out the "by the way," you'd have, "this is why that artificial neurons have..." I'd change that up somehow to avoid the awkward phrasing. Maybe something like, "By the way, this is why artificial neurons have..."

love the nod to Growth Mindset :)

5:00 - "crappy". Not a problem for me personally, but if someone wanted to show this in a classroom, it might raise an eyebrow. "Awful" would work just as well. 8:33 - Insanely huge, or insanely high dimensional?

Max Goldstein

Fantastic! Small point - should 5:06 say "Growth mindset" rather than "Growth midset"?

I'm somewhat qualified to answer this so might be able to help. The basic idea behind a GAN, or generative adversarial network, is that you have two networks competing. One aims to discriminate real images from artificial ones, the other aims to generate genuine looking artificial images which will fool the discriminator. By simultaneously optimising both competing networks you can end up with an extremely impressive generative network. This type of network is just one of many which operate in the realm of 'unsupervised learning' where an external objective is not made explicit. Other interesting examples include Boltzmann machines, Helmholtz machines and variational autoencoders. It would be fantastic if we got videos on some of these although Geoff Hinton also has an excellent course on Coursera.

Sanjeevan Ahilan

I noticed that at 8:28, your negative gradient vector is missing those "vertical continuation" dots. Also, a cool connection to explore might be singular value decomposition, which more literally decomposes the images by their structural dependencies - I recall making a classifier with around 96% accuracy by creating a matrix composed of all the test "0"'s or whatever number, computing the SVD that computes a basis made of vectors that can be explained as the "0"'s singular images and using the sum of the first k or so such images as the "perfect 0". Then all the test vectors that had a euclidean distance closest to that are classified as "0". It might be somewhat offtopic for the series, but would likely make an interesting video.

Awesome. I meant to ask this in the first episode - will you cover how some neural networks can also create new data that matches the original? For example, <a href="https://www.youtube.com/watch?v=p_7GWRup-nQ" rel="nofollow noopener" target="_blank">https://www.youtube.com/watch?v=p_7GWRup-nQ</a> generating human faces based on training data of lots of human faces. I've always vaguely understood the basic concepts of how neural networks work (though it's definitely soooo much better now that you're explaining them), but I don't really understand the leap between learning to identify data and learning to *generate* data.

Love it :) Small typo at 5:03 though

Janik

Right on! Wasn't sure if it was my wifi, use of a copy feature, or whatever. I will give it some time propogate then go back. Thanks.

Bill Russell

You are rather early here, so YouTube might not be done processing.

3blue1brown

Not sure of the reason or cause, but the video is fuzzy not the usual crisp clear picture. I had to stop the video to make this comment. I will go back to it and try to watch the rest.

Bill Russell

At 2:10 the "Guess" text, arrow below it and the square to its left wiggle left and right for no good reason – you should probably not right-align that :)