The timing of action potentials of sensory neurons contains substantial information about the eliciting stimuli. Although computational advantages of spike-timing-based neuronal codes have long been recognized, it is unclear whether and how neurons can learn to read out such representations. We propose a novel biologically plausible supervised synaptic learning rule, the tempotron, enabling neurons to efficiently learn a broad range of decision rules, even when information is embedded in the spatio-temporal structure of spike patterns and not in mean firing rates. We demonstrate the enhanced performance of the tempotron over the rate-based perceptron in reading out spike patterns from retinal ganglion cell populations.
Extending the tempotron to conductance-based voltage kinetics, we show that this model can subserve time-warp invariant processing of afferent spike patterns. Furthermore, we show that the conductance-based tempotron can learn to balance excitation and inhibition to match its integration time constant to the temporal scale of a given processing task. These mechanisms enable already small populations of model neurons to match the performance of state-of-the-art speech recognition systems on isolated word recognition tasks.