Hebbian Learning of Temporal Correlations: Sound Localization in the Barn Owl Auditory System

Until recently an unresolved paradox existed in auditory and electrosensory neural systems [1, 2]: they encode behaviorally relevant signals in the range of a few microseconds with neurons whose time constants are at least one order of magnitude bigger. In this paper we will explain how the paradox can be resolved. We take the barn owl’s auditory system as an example and present results of a modeling study [3] of a neuron in the laminar nucleus that combines inputs from both ears: (i) A single neuron can be quite a good coincidence detector despite its receiving stochastic input [20]. The reason is that it is driven by many presynaptic signals that arrive more or less coherently and that these signals are, to a decent approximation, stochastically independent of each other. (ii) Coherence presupposes the “right” underlying hardware which operates with atcs precision. A simple proposal were to generate the hardware by genetic coding but that is hard to believe since thousands of axons require a immense amount of genes that were not able to adapt themselves to fluctuating circumstances such as food. In our opinion, the way out is provided by an unsupervised Hebbian learning rule that selects synapses and, hence, axons with the “right” delays and suppresses the rest. This is — in a sense — an “evolutionary” process with a “survival of the fittest”. (iii) Combining two groups of input, say from the left and the right ear, one gets the very same tuning. Evaluating the output of many laminar neurons through a population code one arrives at the final µs precision.