The effect of temporal predictability on habituation : empirical studies and connectionist models

This thesis reviews existing models of habituation of the orienting response (OR) and presents four experiments designed to clarify contentious issues in the existing literature. A new connectionist model of habituation is developed and its performance in various experimental situations simulated. Models created to account for habituation can be classified into two types: 1)Comparator models in which the subject compares an internal trace of past stimulation with current stimulation and the orienting response is an index of the disparity between the two. 2) The dual-process model in which habituation is the result of two independent processes of inferred habituation and inferred sensitisation acting within the stimulus-response (S-R) pathway. A major difference between the two types of model is that comparator models of habituation (Ohman, 1979; Sokolov, 1960; Wagner, 1981) postulate that the internal trace encodes extrapolatory or temporally predictive information about the stimulus series, while the dual-process model (Groves & Thompson, 1970) does not suppose that temporal information is encoded. Both the elicitation of a response by stimulus omission, and the effect of 1ST variability on the rate of habituation, are vital indicators of the merit of each approach. Only if both phenomena are demonstrated to exist must a valid habituation model incorporate a mechanism of temporal encoding. Prior evidence for a response to stimulus omission and an effect of 1ST variability on rate of habituation was inconclusive. Four experiments were conducted to examine these and related phenomena in habituation of the skin conductance response (SCR). It was concluded, from the results of these experiments, that both response to stimulus omission and an effect of ISI variability on the course of habituation had a significant effect at short ISIs (1- 2s) but no significant effect was at longer ISIs (>10s). A model of OR habituation was devised that was capable of incorporating the new results as well as other published results. A neural network or - connectionist modelling framework was chosen for this purpose, for its quantitative nature, ease of simulation and neurobiological plausibility. Five different connectionist models were constructed and simulations were performed to assess the performance of each model in experiments testing various temporal phenomena. It was concluded that a model incorporating delta-rule learning of physical features plus learning of temporal relationships by the learning of the interval between stimuli could most satisfactorily simulate the observed empirical results. The connectionist models and their performance in experimental simulations were related to traditional models of habituation