Hippocampus

Hippocampus apparently represents more abstract concepts than even IT

Recurrent final layer (or more complicated neural turing machine or even a transformer) at the top of a feedforward net could be a good enough approximate model? Would be cool to build a model like this.

Open Project ideas
Question - Does the hippocampus develop earlier or later than the cortex or all at once? Would be important if we want to just stick a memory net on top of LPL-like SSL models.

Architecture:

Function:

• CA1 - sequence representation (place cells, time cells etc)
  • NMDA knockout here messes with spatial memory
• CA3 - autoassociative network (Hopfield net)
  • NMDA knockout here messes only with cue-recognition but not with spatial memory
• DG - sequence initiator (the impulse generator)
  • Heavily competitive network - extremely sparse activity
  • Powerful output synapses