A Learning Rule for Extracting Temporal Invariances

Neurons recode information implicit in their inputs. Within a population of neurons, this recoding process is often associated with the removal of redundant information. Clearly, an important objective is to extract a perceptually salient invariance, whilst also specifying its value. A linear combination of Hebbian and anti-Hebbiaa adaptation (HAH learning), operating simultaneously upon the same connection weights but at different time scales, is shown to be sufficient for the unsupervised learning of temporal invariances. A model neuron which implements this rule learns to detect an invariance, and also specifies the value of that invariance.