1_PlaceCells.md

Hippocampal Place Cells (2014 Nobel Prize)

from brain_modules.hippocampus.PlaceCells import (
    HippocampalPlaceCells,
    supervised_learning,
    unsupervised_learning,
)

HippocampalPlaceCells(
  (encoder): Sequential(
    (0): Linear(in_features=1, out_features=128, bias=True)
    (1): ReLU()
    (2): Linear(in_features=128, out_features=128, bias=True)
    (3): ReLU()
    (4): Linear(in_features=128, out_features=16, bias=True)
  )
  (decoder): Sequential(
    (0): Linear(in_features=16, out_features=128, bias=True)
    (1): ReLU()
    (2): Linear(in_features=128, out_features=128, bias=True)
    (3): ReLU()
    (4): Linear(in_features=128, out_features=1, bias=True)
  )
  (predictor): Sequential(
    (0): Linear(in_features=17, out_features=128, bias=True)
    (1): ReLU()
    (2): Linear(in_features=128, out_features=128, bias=True)
    (3): ReLU()
    (4): Linear(in_features=128, out_features=16, bias=True)
  )
)

Supervised Learning

• observation -> MLP -> position embedding (each dim of the embedding is a place cell!)
• The MLP is trained such that place cells respond to observations (positions) with Gaussian
• Note: animals cannot take coordinates as inputs, we only have visual observations, and we learn about our positions by transforming these observations into embeddings (physical coordinates encoded as high dimensional vectors)

Unsupervised Learning (World Model)

• encoder: observation -> MLP -> position embedding

• decoder: position embedding -> MLP -> observation

• predictor: (embedding, action) -> MLP -> next embedding

• Failures:

  • I didn't manage to produce Gaussian response
  • The responses degenerate to linear at one place cell, while others give zero (due to sparsity loss)