DeepDream: How Alexander Mordvintsev excavated the computer’s hidden layers | Hacker News Reader

DeepDream: How Alexander Mordvintsev excavated the computer’s hidden layers

thereader.mitpress.mit.edu

I've been incredibly lucky to work with Alex on several projects, including DeepDream. He's amazing. If you think you have a new idea about how to understand neural networks, there's a decent chance Alex did a prototype of it five years ago.

Regarding DeepDream, it often feels to me -- I don't wish to speak on behalf of Alex or Mike -- that we didn't really understand what our results meant when we published DeepDream. It was kind of like discovering that warped glass can distort and magnify images: a really interesting discovery, but a lot more work was needed to turn it into a scientific instrument like glass can be used to form a microscope. As the community got single neuron or direction feature visualizations that worked well, lots of research possibilities began to open up. And in retrospect, one of the most important tricks was jitter, which Alex introduced. This style of feature visualization is probably the single tool I rely on most in my research to this day.

(If you're curious what this has led to as we've continued to pursue it, check out Circuits (https://distill.pub/2020/circuits/zoom-in/), Building Blocks (https://distill.pub/2018/building-blocks/) and Activation Atlases (https://distill.pub/2019/activation-atlas/).)

I'd also encourage people to check out Alex' new line of research, Neural Cellular Automata (https://distill.pub/2020/growing-ca/). I think it's a really interesting line of exploration. And as usual, Alex has an incredible deep trove of small fascinating results relating to NCA if you talk to him about it.

> The crucial point is that the machine does not see a cat or dog, as we do, but a set of numbers.

This seems to miss the point - to follow that pattern, "Humans do not see a cat or a dog, they receive a set of neural impulses".

If a human "knows" those impulses represent a cat, you could also surely say an artificial neural net "knows" those numbers represent a cat - and if you ask "how" a human/NN knows this, I guess the answer is the same -- different levels of visual abstraction (numbers/impulses trigger neurons that recognize edges and shapes, which become eyes become faces become bodies become animals...) trigger different levels of the network that are familiar with those abstractions and turn them into the end result: "That is a cat."

Ugh, I really dislike articles that are about to tell you the key idea(s) at the beginning then veer off into a personal interest story before doing the reveal. It's enough to get me to quit the article.

To be fair, judging by the title this is indeed a personal interest story. (Edit: It is another thing that way too many popular articles disguise themselves as personal stories as though assuming that more readers will be attracted to a tabloid kind of piece than to the subject itself; the problem may be that most subjects are old and generally not interesting any more, whereas new personal stories appear every day!)

In college I took a reading class for quantum computation .

The thing I learned about reading anything is reading the abstract and the ending and then reading the middle part if you find it interesting.

Ok, but please don't post unsubstantive comments to Hacker News.

It is a substantive comment about the presentation of the subject matter.

I see comments on how articles or websites are presented all the time, didn't realize it was disallowed. Apologies.

Thanks—appreciated.

Reply on news.ycombinator.com