Testing how effective "natural selection" is in producing accurate neural networks

To install and run this project locally, copy and paste the following into your terminal:

git clone https://github.com/giyushino/NeuroEvolution
cd NeuroEvolution
chmod +x /scripts/install.sh
./scripts/install.sh

If you have any trouble, it's likely that you need to install the proper version of Pytorch with GPU support for your device.. I'm using CUDA 12.8.

• Write CNN, ViT (pytorch + jax for both)
  • Putting jax off for now, not exactly sure how to speed up model initialization, which is necessary for evolutionary stuff
• Set up 2 datasets
  • Google Doodle + real images -> i'll do real later
• Test if parsing the raw jsonl is faster than datasets (i think it probably is)
  • Doesn't matter I think, both are fast enough
• Rank the classes by how similar they are and see if we can get the ordering correct, ie dragon and crocodile should be similar to each other
  • Maybe save embeddings
• Set up normal training pipeline
  • know it works for cnn and linear, idk about vit yet
• Train all 4 models
• Create function to compare weights (cosine similarity)
• Write evolutionary algorithm
  • for some reason converging super slowly, write it again
• Test model robustness (add noise)
• Create dataset with predators, see if model learns to differentiate between different animals
  • Like if we know ducks don't hurt us but crocodiles + lions will, is there any point to learning the difference between croc/lion
  • study the embeddings ig
• See what features are the most important
  • Grad Cam, see if this allows us to determine what layers mutations should affect the most
  • hooks, check out some other mechanistic interpretability techniques
• Plot loss landscape
• Front end maybe?
• Animations?