/ TeLC

  • destroys some synatic protein
  • blocks release of neurotransmitter, but neuron still spikes
  • nice way to test if feedback is necessary or not or just lower order recurrence is enough

haim

  • if N dims, then max num of linearly seperable data pts is 2 N alpha = num of manifolds / num of neurons

seperable manifolds

  • tangled = 2/m untangled
  • ~ 1 R_m sqrt(D_m) </> 1 for untangled/tangled

he is not putting anythingm its already pre-trained. each layer, linear decodng, that i what he is telling

but he is missing hierarchy no

language to vision

what is special about the mapping? could any mapping would have done it, would not mappin gwould have done it? possible good representations that could be mapped to. resnet last layer need not be the best represnetation?

  • … language aligns not with arbitrary visual statistics, but increasingly with representations that encode predictable and semantic object structure.