Growing neural cellular automata in PyTorch

In this video, I implement the Growing Neural Cellular Automata article. It is a biologically inspired deep learning pipeline that generates update rules that are applied to a grid of pixels. It uses heavily the convolution operation together with multiple other techniques - alive masking and stochastic update.

00:00 Intro + resources
01:10 Motivation and setup: Game of life
04:12 Learning rules using convolutions
06:04 Article: proposed pipeline
08:35 Model: constructor
11:35 1x1 convolution explained
13:29 Model: constructor continued
14:29 Model: helper methods
17:22 Model: forward pass
18:43 Training script
24:52 Launch training
25:20 Results
25:44 What was not covered
26:01 Outro

Credits logo animation

[Larmbs]

You are incredible man.
-You go at a good pace.
-Each project feels well planed.
-Nice formating style.
-Good explanation.

Ive just started really digging into this machine learning space, any recommendation on learning on all the different layer types, and problem types?

[nabeelhasan]

Your channel is goldmine, it took a year for YouTube to recommend this gem so happy to find your work

[futurisold]

Just found you, and I couldn't be happier. I've seen that whenever you're manipulating tensors, you're trying to show the impact of the transformation. Please keep doing so! Furthermore, it would come in handy if you would explain in greater detail the dimensions of the tensors and why certain operations are required (I mean, I would keep this as a rule because otherwise, it might get hard to follow). Looking forward to seeing your future uploads!

[hasanhuseyinyurdagul]

Best youtuber the universe has ever seen

[alexanderk]

hey, thank you for the video, really great job on explaining the original parer architecture

[Mome]

Hey ! Thanks dude, I really enjoy this format ! ((:

[mrigankanath]

Awesome!! and a very good topic.
Still waiting on your video on Attention layers😝

[danassouline]

Hey thanks for the video and the code that's very nice! Quick question: I tried to re-use your code, which indeed works but for some reason the eval video doesn't seem to appear in tensorboard after the training is complete (the target image does appear and the loss does show up normally) - I know Tensorboard fairly well but i've never exported a video on the GUI before: is there something else you need to do in addition to adding the writer.add_video() line? Thanks a lot!