Walter Senn, University of Bern
Learning by the dendritic prediction of somatic firing
I will present a biological form of learning in a dendritic tree with conductance-based synapses. The neuron is considered as an intrinsic prediction element that receives inputs from two sources of information: “teaching afferents" targeting the somatic region, and “student afferents” targeting the dendritic tree. The synaptic weights on the dendritic tree are adapted such that the local dendritic potential predicts the somatic firing, evaluated in terms of the back-propagating action potential. I show that this view of learning is compatible with most of the spike-timing plasticity experiments. It also yields an unified framework for supervised, unsupervised and reinforcement learning. I will introduce the notion of prospective coding according to which a neuron learns to predict future somatic inputs. When the somatic input encodes reward, the neuron learns to predict the discounted future reward. In this case, the learning scheme yields a neuronal implementation of TD(lambda). Another application
refers to the learning of neuronal integrators.