Delighted @jajmca.bsky.social's abstract was accepted for COSYNE 2025, for his work on linking neural circuit structure to function via drosophila-connectome-based reservoir computing. With @conjh.bsky.social, John Wade and myself.
Comments
Log in with your Bluesky account to leave a comment
I'm a big fan of that paper! My guess is that for training internal RNN weights these days most people use backpropagation through time, or similar gradient-based methods. But you still see FORCE used a bit
Thanks! Do you have a sense of how FORCE and in general that type of approach (leverage innate trajectory) performs vs back prop and other uses of RRNs. Do you think there is any opportunity in exploring it from the perspective of efficiency or scalability?
Comments
Your SBM drosophila paper was one of the direct inspirations for this work btw, so thanks!
Specifically, I’m curious how common use of ‘innate trajectories’ are before then trining output weights?
Laje Buonomano 2013
https://pubmed.ncbi.nlm.nih.gov/23708144/