The SP tiebreaker is necessary to enforce strict sparsity control, but it has undesirable sideffects in other properties:
sp(z) becomes stochastic (random seed from hash(z)?)
noise robustness suffers and doesn't improve a lot with training
This points to a need to rethink inhibition, and not do it by comparing against a threshold, but more low-level,
by ordering indices and picking the highest activations.
The SP tiebreaker is necessary to enforce strict sparsity control, but it has undesirable sideffects in other properties:
sp(z)
becomes stochastic (random seed from hash(z)?)This points to a need to rethink inhibition, and not do it by comparing against a threshold, but more low-level, by ordering indices and picking the highest activations.