Depends on how the prediction is measured. If the model units that are being compared with (e.g.) V1 are only the final layer of the model, then the hidden activities and internal mechanisms of the model are massively underconstrained.
Comments
Log in with your Bluesky account to leave a comment
Also, using a model architecture that violates a lot of known structural details of cortex (e.g. Dale's law; recurrence) seems like a recipe for not learning correct mechanistic details.
Comments