> It randomly inserts layers of computation between the networks’ input and output, and tests and trains the resulting wiring with the limited data. If an added

> It randomly inserts layers of computation between the networks’ input and output, and tests and trains the resulting wiring with the limited data. If an added layer enhances performance, it’s more likely to be included in future variations. Sounds a lot like genetic algorithms but with neural networks. I suspect we'll see more of this as people figure out how to run the search over neural network architectures that fit their own domains. Convolutions and transformers are great and all but we might as well let the computers do the search and optimization as well instead of waiting on human insights for stacking functions.