Stability versus Neuronal Specialization for STDP: Long-Tail Weight Distributions Solve the Dilemma

Abstract:

Spike timing-dependent plasticity (STDP) is hypothesised to structure neuronal networks in the brain depending on their precise spiking activity. In this talk, we focus on the functional implications of an experimentally observed property of STDP, the dependence of the weight update on the current strength of the synaptic weight. This weight dependence crucially shapes the distribution of plastic synaptic strengths. We propose a model of weight-dependent STDP that can generate long-tail (e.g., lognormal) distribution of synaptic strengths that have been observed in recent experiments. We show that our new model offers a balance between stability for the weight distribution and competition induced by input spike-time correlations, taking the "best" of additive and multiplicative STDP. This allows a neuron to robustly and quickly adapt to changes in the structure of input spike trains, e.g., in a context of short-term memory for example. In a recurrently connected network, our new model allows the persistence of a stable weight structure susceptible to represent the structure of input spike trains.