Neuronal dynamics
For a single neuronal layer behavior is stochastic in absence of stimuli, periodic and largely phase synchronized with presence.
Train bottom up pathways to target post-lateral influence values while top-down targets pre-lateral influence values
Instead of training N parameters on S steps of gradient descent slowly, train parameters individually for N steps. O(NS) > O(N)
In equilibrium,
B ≈ N ≈ S
Weight updates per batch O(1) instead of O(n)
Each of N gradient descent updates costs O(n). Total cost O(n^2)
With single most active assignment, only O(1) computations are run per optimization step. Total cost O(n)
Forward pass still costs