Spike-driven synaptic plasticity for learning complex patterns
of mean firing rates
Stefano Fusi, Joseph Brader, Walter Senn
Spike-driven long term synaptic plasticity has been investigated in
simplified situations in which the patterns of mean rates to be encoded
were statistically independent. A regulatory mechanism is required to
extend the learning capability to more complex and natural stimuli. We
propose a spike-driven synaptic dynamics which combines the depedence on
spike timing with the depedence on the post-synaptic depolarization. The
learning rule implemented by these dynamics produces long term changes
when the pre-synaptic neuron is active and the post-synaptic mean firing
rate is in a limited range. The fact that the post-synaptic activity
becomes too high or too low is interpreted as a signal that the
post-synaptic neuron is already responding as required by the activity
imposed during training and learning stops. The acquired information about
the neural activity is preserved on long time scales by the inherent
bistability of the synaptic dynamics. The bistability allows also to
modify a small fraction of synapses upon each stimulus presentation,
randomly selected if the neural activity is irregular. This selection
mechanism allows to recover the optimal storage capacity of the network
even when the synapses can preserve only two states on long time scales.