Fig. 4: Tradeoff between time and dynamic range of synaptic currents over which contrastive learning is approximated.
From: Temporal Contrastive Learning through implicit non-equilibrium memory
A The sawtooth training protocol sij(t) with amplitude A approximates contrastive learning in specific regimes of protocol timescales (τs, τf) relative to nonlinearity threshold θg and synaptic memory kernel timescale τK. B The weight update Δwij is proportional to the signal amplitude sclamped − sfree as desired within a range of amplitudes (regime ii) for given timescales. C Learning range (defined in (B) as the dynamic range Amax/Amin for which the sawtooth protocol approximates contrastive Hebbian learning) plotted against τs/τf. Signal timescales (τs, τf) place a limit on maximal learning range which increases with increasing τs/τf. Protocols with higher τf/τK approach this limit.
