Opposition logic and neural network models in artificial grammar learning☆
Introduction
Tunney and Shanks (2003) responded to our experiments in which we used opposition logic to separate controlled from automatic influences in artificial grammar learning (Higham, Vokey, & Pritchard, 2000). They used a simple neural network simulation that they claimed produced similar results without the need of separate processes or influences. There is a somewhat dismissive tone in Tunney and Shanks (2003) that seems to imply that a major contention that we made has been disproven. We are not sure what claim we have made that has been disproven by their simulation, but we do think that the value of the experiments reported by Higham et al. (2000) is being underestimated in their discussion.
Tunney and Shanks (2003) argued for a single memory system, which we also have consistently argued for over at least the last 10 years (see, e.g., the “episodes and abstractions” section of Higham et al., 2000). This claim is not a matter of contention between us, and, indeed, is rather a growing consensus in the literature (e.g., Banks, 2000; Palmeri & Flanery, 2002; Ratcliff, Van Zandt, & McKoon, 1995). What appears to be at contention is whether the opposition procedure provides evidence that is usefully interpreted as a distinction between controlled and automatic influences. Manipulations intended to vary the degree of participants’ control, such as deadlines (e.g., Higham et al., 2000; Toth, 1996), divided attention (e.g., Debner & Jacoby, 1994; Jacoby, 1991), or delayed testing (e.g., Higham et al., 2000; Jacoby, Kelley, Brown, & Jasechko, 1989a), produce separable effects when investigated with the opposition procedure. Whether this separation of effects is called a difference in influences or processes or changes of parameters within a single system, the opposition procedure has produced clear and stable patterns of results in research on memory, word perception, and artificial grammar learning. Over a wide range of experimental variations, the effects of these variables are computationally separable into controlled and automatic components, both of which may be manipulated without influencing the other: deadlines and divided attention selectively affect the controlled component (e.g., Toth, 1996; Yonelinas & Jacoby, 1994), whereas variations in such things as habit strength and proactive interference do so for the automatic component (Hay & Jacoby, 1996; Jacoby, Debner, & Hay, 2001). Furthermore, this pattern of influence is clear and simple using opposition logic, as opposed to its traditional alternatives, which was the central contrast of our paper.
In what follows, we briefly review the critical results of Higham et al. (2000), and Tunney and Shanks’ neural network simulation of them. We demonstrate that their model failed to reproduce the central pattern of dissociations from Higham et al. (2000). We present a different neural network simulation that provides a better fit with the results, and, finally, discuss the implications of such models for opposition logic.
Section snippets
Higham, Vokey, and Pritchard (2000)
Higham et al. (2000) conducted two artificial grammar experiments during which participants first learned training items all of which conformed to a finite state grammar. Unlike most artificial grammar experiments (e.g., Higham’, 1997a, Higham’, 1997b; Reber’, 1967, Reber’, 1989; Vokey & Brooks, 1992), two different grammars were used to construct the training stimuli, henceforth referred to as Grammar A (GA) and Grammar B (GB), and participants studied the strings under headings that
The Tunney and Shanks (2003) model
Tunney and Shanks (2003) simulated the two experiments on the learning of artificial grammars of Higham et al. (2000) using cosine similarities derived from the learning of the study items in a variant of a serial, recurrent network connectionist model (Elman, 1988), often referred to as a “simple recurrent network,” or SRN. It was first proposed as a model of the learning of single artificial grammars by Servan-Schreiber, Cleeremans, and McClelland (1989) (see also Cleeremans, 1993; Cleeremans
The autoassociative network model
The failure of Tunney and Shanks’ (2003) neural network simulation to reproduce the critical dissociations of Higham et al. (2000) motivated us to investigate whether another neural network simulation could capture these patterns. Our simulations of the results of the two experiments of Higham et al. (2000) used a linear autoassociative neural network. The hidden units of autoassociators are in the form of the eigenvectors of the covariance space of the input stimuli, which can be solved for
Discussion
What conclusions are we to draw from our successful simulations of the experiments in Higham et al. (2000)? Tunney and Shanks (2003) claimed that their simulation of Higham et al.’s (2000) results with a similarity-based neural network model necessarily undermined the logic of opposition because, in their view, such simulations implied only a single system. On the other hand, we would argue that even such simple systems are susceptible to multiple influences, some of which the system can
References (36)
Connectionist and memory array models of artificial grammar learning
Cognitive Science
(1992)A process dissociation framework: Separating automatic from intentional uses of memory
Journal of Memory and Language
(1991)- et al.
Memory systems and perceptual categorization
Implicit learning of artificial grammars
Journal of Verbal Learning and Verbal Behaviour
(1967)- et al.
Analogy and abstraction strategies in synthetic grammar learning: A functionalist interpretation
Cognition
(1978) - et al.
Does opposition logic provide evidence for conscious and unconscious processes in artificial grammar learning?
Consciousness and Cognition
(2003) - et al.
Connectionist models of face processing: A survey
Pattern Recognition
(1994) - et al.
Neural networks
(1999) - et al.
More about the difference between men and women: Evidence from linear neural networks and the principal component approach
Perception
(1995) An introduction to neural networks
(1995)
Recognition and source memory as multivariate decision processes
Psychological Science
Mechanisms of implicit learning: Connectionist models of sequence processing
Learning the structure of event sequences
Journal of Experimental Psychology: General
Unconscious perception: Attention, awareness, and control
Journal of Experimental Psychology: Learning, Memory, and Cognition
A case of syntactical learning and judgment: How conscious and how abstract?
Journal of Experimental Psychology: General
Separating habit and recollection: Memory slips, process dissociations, and probability matching
Journal of Experimental Psychology: Learning, Memory, and Cognition
Chunks are not enough: The insufficiency of feature frequency-based explanations of artificial grammar learning
Canadian Journal of Experimental Psychology
Cited by (0)
- ☆
Preparation of this work was supported by a Natural Sciences and Engineering Research Council of Canada operating grant to J.R.V. and a British Academy research grant to P.A.H. We thank Professor Lee Brooks for his very helpful comments on an earlier draft of this paper.