I'm at Cognitive and Computational Neuroscience 2023 (https://2023.ccneuro.org) in Oxford , and enjoying it.
Most of the time I work on synaptic and cellular stuff and a bit of neural population data analysis. So the cognitive and ML/AI angles at this meeting are out of my comfort zone, but also the reason I'm here... much easier to get up to speed on the gist of a research area by spending a few days at a meeting than trying to decipher what's going on reading papers at home on your laptop
A "generative adversarial collaboration" session yesterday on "Comparing artificial and biological networks: are we limited by tools, hypotheses or data?" https://gac.ccneuro.org/gacs-by-year/2023-gacs/2023-2
Some cool stuff but overall I found the session unsatisfying:
- many speakers crammed a 15-min talk into their 8 min slot and advertised their work instead of making a point
- moderators didn't police time which meant audience questions (20+ good ones on website) got cut
- no general consensus beyond trivial points
Comparing artificial and biological networks: are we limited by tools, hypotheses or data? Organizers & Speakers at CCN 2023: Meenakshi Khosla, Massachusetts Institute of Technology Apurva Ratan Murty, Massachusetts Institute of Technology Tal Golan, Ben-Gurion University of the Negev Jenelle
So are ANNs valid or useful as models of the brain?
Absolutely, but - like all models - only for a subset of questions, and only if used in a targeted way. One of my PhD students James Malkin is using them in this way and his project has turned out great.
But imo they are not on their own "the answer", and will not completely displace classic physics-style simple models.
One exciting hybrid area to me is using gradient descent etc to fit biologically plausible models to perform tasks
Cian O’Donnell