Rule Learning

Dataset size has only mild effects on τrule in the parity experiments. τrule is defined as the first step where sample-level rule accuracy reliably exceeds 0.9. Increasing parity group size delayed rule learning by roughly an order of magnitude for G=2 through G=4. Rule learning is controlled mainly by rule complexity. Rule learning corresponds to expansion of attractor basins for rule-valid vertices and shrinkage of rule-violating basins. DSM loss separation for rule learning occurs mainly at intermediate noise scales.