Mechanisms of implicit learning: a parallel distributed processing model of sequence acquisition
Axel Cleeremans · 1991
In this thesis, I examine implicit learning--the process by which knowledge about the rule-governed complexities of the stimulus environment are acquired independently of conscious attempts to do so (Reber, 1989)--from an information processing perspective. I argue that most work on implicit learning has been distinctly atheoretical, and that progress in this field requires mechanistic models of the underlying processes to be developed. I first review experimental evidence and recent information processing models of implicit learning. I propose a general framework for thinking about implicit learning processes, based on a series of principles constrained by theoretical and empirical considerations. Next, I present a detailed PDP model (the model) of implicit learning performance in a complex sequence acquisition paradigm, and two experiments exploring subject's ability to encode temporal information in this paradigm. The results indicate that subjects become progressively sensitive to the temporal structure of the material, despite their being unaware of the relevant contingencies. The model is successful both in accounting for human behavior and in instantiating general processing principles characterizing implicit learning mechanisms. I then explore how this model and the empirical work may be extended to cover additional phenomena and theoretical issues pertaining to sequence learning. I report on simulations of the effects of attention and of explicit knowledge on sequence learning performance, describe the performance of an amnesic patient, and explore how the SRN model may be applied to explicit prediction tasks. Together, this set of simulations provides additional support for the model. In a third experiment and accompanying simulations, I explore how well subjects can maintain information about remote context. The results do not fully support the SRN model: It appears that even simpler, decay-based mechanisms, may successfully account for the data. I discuss the implications of this finding for further research. The Appendix contains a reprint of a paper by Cleeremans, Servan-Schreiber and McClelland, in which we report on the computational characteristics of the SRN model. We show that the SRN is an instantiation of a new class of computational objects that we call Graded State Machines.