Fig. 3

The presynaptic terminal adapts to the statistics of pre- and postsynaptic activity. The presynaptic learning rule can be understood as a minimisation of the prediction error between neurotransmitter release and postsynaptic spiking. a Hebbian activity, whereby postsynaptic spiking (in the form of somatic or dendritic spikes) is causally paired with presynaptic action potentials triggers an increase in Pr by a positive feedback signal (i.e. retrograde NO signalling). b Release of glutamate causes a decrease in Pr irrespective of postsynaptic spiking by a negative feedback signal (i.e. presynaptic NMDAR activation). c Positive and negative feedback signals work in parallel and cancel each other out when neurotransmitter release is followed by postsynaptic spiking. d This learning rule optimises Pr with respect to the conditional probability of postsynaptic spiking given prior presynaptic activity. At steady-state, each release event will be followed, on average, by a postsynaptic spiking event (Left). As a consequence, burst firing will generally result in low Pr (Middle), whereas synapses using spike-timing codes will tend towards high Pr (Right). Pr is therefore optimally tuned to preferentially transmit presynaptic input that is predictive of postsynaptic spiking