Math, Neuroscience, ML. Postdoc at Imperial College London; also writes music and dreams of post-growth economies 🧠🪴🇵🇸 Website: https://computationalcognition.ca/author/
Ezekiel Williams
This Korean paper asmr artist that I’ve been following for a while made a TikTok of her healing Van Gogh 🥹
Her account name is paper pepper
www.tiktok.com/@paper__pepp...
6/7 Lastly, we also show that certain local learning algorithms appear to be restricted in the rank of solutions they are capable of learning. We prove this mathematically for RFLO in linear RNNs, generally (no data-aligned assumptions required).
Video
Oups, forgot to hyperlink to @glajoie.bsky.social's account in the first post. Will also take this moment to thank him for being a great PhD supervisor!! (thank you Guillaume!)
2/7 Synaptic weight changes in local learning models depend only on information proximal to a synapse in space and time. This is often required in models of how the brain learns and in neuromorphic methods. Our results should thus be of interest to neuroscience and machine learning researchers.
5/7 We then numerically validate this theory, showing that our analyses provide insights even when networks do not fully satisfy the restrictive data-alignment assumptions.
