Neuro background here. There's a lot of overlap, honestly. Polysemanticity is what we call mixed selectivity, and the CLIP multimodal neurons paper directly compares its findings to "concept cells" like the Jennifer Aniston neuron.
Also a couple analogies:
Feature viz <=> receptive field mapping
patching/ablation <=> lesion/optogenetics
The big difference is access. Interp is easier to intervene, neuro is stuck inferring from a handful of noisy electrodes with a bunch of what I call evolutionary legacy code.
See: Jonas & Kording's "Could a Neu... (read more)
Neuro background here. There's a lot of overlap, honestly. Polysemanticity is what we call mixed selectivity, and the CLIP multimodal neurons paper directly compares its findings to "concept cells" like the Jennifer Aniston neuron.
Also a couple analogies:
Feature viz <=> receptive field mapping
patching/ablation <=> lesion/optogenetics
The big difference is access. Interp is easier to intervene, neuro is stuck inferring from a handful of noisy electrodes with a bunch of what I call evolutionary legacy code.
See: Jonas & Kording's "Could a Neu... (read more)