Imprinting and recognition memory: a neural net model

Abstract. The neural processes underlying filial imprinting in birds are divided, for purposes of theoretical analysis, into three systems: Analysis of stimuli into features, Recognition of familiar stimuli in terms of their features and Execution of filial behaviour. The Analysis system is linked to the Recognition system which is linked, in turn, to the Executive system; the Analysis system is also linked directly to the Executive system. Each system consists of modules that contain units. The neural net, used to explore this model, is unsupervised and three-layered; and it is fast since it can operate efficiently with few modules. In the naive state, the strengths of the connections between units in one system and those in others downstream are all low, with the exception of those connections between units in the Analysis and Executive systems; these connections are started at maximum strength. Connections between active units of two systems are strengthened up to a maximum; those between an inactive unit of one system and an active unit in a system downstream are weakened down to zero. Weakening takes place more slowly than strengthening. This model simulates many aspects of imprinting including recognition of fine detail in the familiar stimulus, while generalizing to stimuli resembling the one with which the bird was imprinted. When experienced, the model simulates withdrawal from objects that are sufficiently novel and, as a consequence, the end of the sensitive period for imprinting. The model readily forms polymorphous categories, updates existing stimulus representations and classifies together physically distinct stimuli, if they have occurred within a short time of each other.