What’s Inside a Neural Network?

The shown blog post is available here:

📝 The paper "Zoom In: An Introduction to Circuits" is available here:

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Alex Haro, Alex Paden, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bruno Mikuš, Bryan Learn, Christian Ahlin, Daniel Hasegan, Eric Haddad, Eric Martel, Javier Bustamante, Lorin Atzberger, Lukas Biewald, Marcin Dukaczewski, Michael Albrecht, Nader S., Owen Campbell-Moore, Rob Rowe, Robin Graham, Steef, Sunil Kim, Taras Bobrovytsky, Thomas Krcmar, Torsten Reil, Tybie Fitzhugh

Károly Zsolnai-Fehér's links:

#ai #machinelearning

[DrMohamadZeina]

There is probably no journal that would make me as happy to be published in. Great work Dr Karoly, keep it up

[Tinfoilpain]

I dropped so many of my papers.

What excites me about neural networks is that we can inspect the internals - which can give us such great insight to our own minds.

[altbert]

A "ten minutes paper" channel will be nice too

[TaarLps]

Its interesting how much these things look like psychedelic hallucinations.

[avi]

This inspired me to read the whole paper and was delighted by how accessible and philosophical it was. The most interesting idea was that studying features in artificial neural networks could help us understand features detected by the human brain.

[Mnnvint]

Chris Olah is a rare talent for explaining things. His old blog posts on backpropagation and LSTMs are terrific.

[SpaghettiniFiveMillion]

Is there an explanation why almost all neural network generated images look how they look? They all have bright colors, similar patterns etc.

[minidreschi]

2:43 a squirrel and also a fox second in the bottom line :)

[sudhanvagokhale]

@Two Minute Papers glad to have people like you sharing knowledge to all. Keep it up!

[wileyr]

This is actually a really good way to begin to understand basic neural networks. It show how you don’t have ultra defined data sets, you just mash everything together and modify it until it works

[Baleur]

0:45 this seriously looks like some DMT dreamstate stuff...
Holy shit.... I mean okay i know we humans created and tuned these visualizations to look like this for our own amusement (not the patterns, but at least we had SOME input into the color ranges etc so the results wouldnt just be a white incomprehensible noise).
But it is still astonishing how, real human beings, with real conciousness (presumably?), have seen THESE kind of patterns during psychadelic trips.. I mean, our own minds have created the same patterns that we now see neural AI networks use as they're trying to filter the world?
That Really fucked up xD xD

[ThatBulgarian]

is there something stopping these neurons being printed onto canvases and sold as art?

[Kknewkles]

The neural network-generated images that go bad must be the most realistic manifestation of what it would look like to see some eldritch horror stuff, on the very edge of our comprehension. Looks barely recognizable, and at the same time, utterly alien.

Also the three dislikes are the AI neurons misfiring.

[markmarti]

Sajnos aránylag kevés Magyar tud ennyire jól bármiféle idegen nyelvet :( Nagyon jók a videók keep it up <3

[iau]

3:17 "Who devotes his time away from research...". It's so interesting to me that developing ways to make these concepts easier to understand for the average person is so low in priority on researcher's mind.

I'd argue that getting people to understand this stuff is maybe as important as discovering it in the first place.

Have you ever tried to read a scientific paper? A single page sometimes takes hours or even days to understand. Yet when you finally get it you can easily explain it in minutes.

It's almost as if the scientific community puts value on being as convoluted as possible. I'm glad channels like Two Minute Papers exist. If it weren't for you, I would never even begin to have an idea about any of this.

[Theoddert]

Clearly that was not a dog detector but, as you said, a Good Boy Detector. Squirils can be good bois also

[malyalaA]

W&B is a life saver, it really is as good!

[splytrz]

I'm don't understand where they took these images from. Are they just convolutional filters?

[elliotjaffe]

I wonder if there's some way to take this same idea of a single neuron depicting several orientations of an object, like with the dog, and apply it to uv maps / 3d object maps for recognizing 3d structures? These images remind me a lot of some of the crazy 2d texture files that are produced for 3d objects

[memegazer]

Just a reminder if you can't donate at least disable your ad block to support the channel.