Rutgers Symposium on Learning I

 Learning: Association or computation?
(November 4, 1998)
 

Abstracts:
 



 

Robert A. Rescorla (University of Pennsylvania)
The survival of the association

The notion of an association has survived many attacks to remain for centuries a primary theoretical construct in psychology. One reason for this survival is its undergoing important evolutionary changes. As the available data and the general scientific environment have changed, the notion of an association
has adapted with changes in the elements that it takes as its arguments, in the conditions under which it is formed, and in the way in which is exhibited in behavior.  Moreover, the association has survived by increasingly constraining the range of psychological phenomena it claims to explain. As a result of these adaptive changes, the association has continued to be a powerful explanatory and heuristic device, especially as applied to elementary learning processes in animals.
 

C. R. Gallistel (UCLA and Rutgers)
Contrasting conceptual frameworks for the understanding of conditioning: Neurobiological
Implications

There are at least two broad conceptual frameworks within which one can think about the psychological and neurobiological processes that mediate conditioned behavior. One is the associative framework in which conditioned behavior is a consequence of changes in associative strengths (synaptic conductances). The other is a decision theoretic framework in which conditioned behavior is the result of decisions based on comparisons of remembered and currently elapsing temporal
intervals. The talk reviews the differences in the fundamental assumptions made by the two frameworks, in the answers they provide to common questions about conditioned behavior, and in
their neurobiological implications.
 

J. E. R. Staddon (Duke University)
Time and memory: Towards a pacemaker-free theory of interval timing

A popular view of interval timing in animals is that it is driven by a discrete pacemaker-accumulator (PA) mechanism that yields a linear scale for encoded time.  But PA mechanisms are fundamentally at odds with the Weber-law property of interval timing and experiments supporting linear encoded time can be interpreted in other ways.  I argue that the dominant PA theory, scalar expectancy theory (SET), fails to explain some basic properties of operant behavior on interval-timing procedures and
can only accommodate a number of discrepancies by modifications and elaborations that raise questions about the entire theory.  I propose an alternative that is based on principles of memory dynamics derived from the multiple-time-scale (MTS) model of habituation.  The MTS timing model can account for data from a wide variety of time-related experiments: proportional and Weber-law temporal discrimination, transient as well as persistent effects of reinforcement omission and reinforcement magnitude, bisection, the discrimination of relative as well as absolute duration, the choose-short effect and its analogue in number-discrimination experiments.  Resemblances between timing and counting are an automatic consequence of the model.  Transient and persistent effects of drugs on time estimates can be interpreted as well within MTS theory as in SET.  Recent real-time physiological data conform in surprising detail to the assumptions of the MTS habituation model.
 

R.M. Church (Brown)
A Turing test of computational and association theories

Scalar Timing Theory is an example of a well-developed computational theory of timing that involves addition, multiplication, division, and sampling. The Behavioral Theory of Timing is an example of a well-developed association theory of timing that involves state transitions and strengthening of connections. A Turing Test is proposed for the evaluation of current models, and to guide theoretical
developments.
 

Peter Killeen (Arizona State University)
Metonymic Psychology. Why Johnny can't add, brains can't think, andcomputers can't compute

The category error involves attributing aspects of a (hierarchically organized) network to sub- or supra-levels. Such metonymy may be poetic, concise, and colorful. It may also be confusing to the tyro or granting agent, who may search for (or fund a search for) properties that can exist only poetically in the terrain searched. Metaphors can transcend levels, but must eventually fail--a warlike society may comprise warlike individuals, but those are seldom claimed to embody warlike dopamine receptors. In this rumination I shall suggest some operating principles for heritability of metaphors, with special attention to association and computation.
 

J. Gibbon and C. Malapani (Columbia)
Separable storage and retrieval distortions in memory for time in Parkinson's disease

Diseases of the Basal Ganglia and the brain nuclei interconnected with them lead to disturbances of movement and cognition exemplified in Parkinson's Disease (PD) and Huntington's Disease, including disordered timing of movements and impaired time estimation.  Distortions appear in both the variability of time estimation and the location (median) of the distribution of estimates. Both these distortions depend upon dopaminergic regulation in the striatum.  Experimental work with PD patients will be presented showing a dissociation between deficits
in encoding (writing to memory) and decoding (reading from memory) of learned temporal durations. When durations are stored in memory the encoding process is slowed, resulting in overestimates.  In contrast an interference or coupling occurs between remembered time intervals on decoding, such that
the shorter of two remembered durations is overestimated, while the longer is underestimated.  This dissociation suggests separable functional pathways for memory storage and retrieval in neural networks of the human striatum.
 

Anthony Dickinson (Cambridge University, UK)
Causal learning: association or computation?

The acquisition of causal knowledge represents a test case for the contrast between associative (e.g. Shanks & Dickinson, 1987) and computational (e.g. Cheng, 1997) theories of learning. The most problematic phenomenon for associative theory is that of retrospective revaluation. For example, if a compound of a treatment cue (A) and a target cue (B) is a paired with an outcome (AB+), the causal status of the target cue B can be enhanced by subsequent presentations of treatment cue A alone  (A-).
Retrospective revaluation challenges the assumption that the associative strength of a cue changes only on learning episodes when that cue is present. Dickinson and Burke (1996) argued, however, that retrospective revaluation is mediated by within-compound associations. Further evidence will be
presented for the role of within-compound associations, although I shall argue that retrospective revaluation involves separate excitatory and inhibitory associative processes rather than a general error-correcting learning algorithm.

R. Miller (SUNY)
Associations and hat racks: Sometimes useful, but not essential

Most associative models can be rephrased to omit mention of associations, which are intervening variables.  Such rephrased models, if quantitative, are computational models, which I define as lacking cognitive constructs.  An associative model is one of many possible forms of cognitive modeling. Cognitive models provide a heuristic for conceptualizing what is learned and how it is learned at the molar level. Such a heuristic has benefits (and costs) in terms of stimulating new research. Computational models that identify computational steps with physiological processes (and/or anatomical locations) are step-wise testable; hence, they suggest new experiments although different ones than cognitive models.  However, computational models that fail to identify testable steps in processing, closely resemble radical behaviorism and curve fitting, which have their own, very different goals.