GEM: Spatial pooling!, Inhibition, boosting, locality in year 1990

Spatial pooling!, Inhibition, boosting, locality

Forming sparse representations by local anti-Hebbian learning

New TERM: Selectivity vs Sparseness


This paper was the basis to the LaterPooler experiment: htmresearch/projects/lateral_pooler at master · numenta/htmresearch · GitHub

1 Like

I think it’s interesting that (Foldiak 90)'s model had excitatory neurons directly inhibiting each other, instead of using a layer of inhibitory neurons.

The result is that sometimes the input is represented by a single cell. Foldiak refers to this as “local” encoding. Another term is “grandmother” cell.

This demonstrates the role of inhibitory neurons in producing the right level of sparsity, in the “Goldilocks” range in between too dense or too sparse.