Multiple Deep Learning Systems

A second refinement of the original CLS model concerns how to understand the neocortical deep learning system. McClelland et al.

(1995) treat the deep learner as a single network. They recognize, however, this is an expository convenience and that it does not reflect the actual organization of the neocortical system: "We view the neocortex as a collection of partially overlapping processing systems; for simplicity of reference, however, we refer to these systems collectively as the neocortical processing system” (p. 422-423). For our purposes, however, we want to be more specific about the kinds of relatively distinct neocortical deep learners. There is now substantial evidence—drawn from studies of differential developmental trajectories during youth, studies of the effects of localized brain lesions, comparative studies across animals, and studies using neuroimaging—that humans have a number of partially dissociable "conceptual learning systems” (Carruthers, 2006). A list of these systems should at least include the following four categories, with multiple subsystems grouped under each.

Statistical learning systems. These systems maintain and update degrees of belief in prospects and calculate confidence bounds on these probabilities (White, Engen, Sorensen, Overgaard, & Shergill, 2014). They also compute decision-guiding signals, including riskiness of outcomes, absolute and relative value, expected value, and discrepancies between expected and actual outcomes (Montague & Berns, 2002; Montague, King-Casas, & Cohen, 2006; Schultz, 2000).

Causal learning systems. These systems generate graphical maps of a domain with links representing causal relations (Gopnik & Glymour, 2002; Gopnik et al., 2004). The graphical causal maps are produced by tracking statistical information about correlations and conditional probabilities among observed events as well as information gained through active interventions (e.g., pushing a button on a machine to see what happens next).

Analogical reasoning systems. These systems apply strategies and models learned in one domain to other domains. This is made possible by identifying abstract similarities at the level of the relations between objects, even when the relevant objects are dissimilar (Gentner & Markman, 1997; Holyoak & Thagard, 1996).

Social cognition systems. These systems keep track of other peoples' mental states as well as underlying character traits that give rise to long-term patterns of behavior (Brune & Brune-Cohrs, 2006; Gallagher & Frith, 2003; Karmiloff-Smith, Klima, Bellugi, Grant, & Baron-Cohen, 1995). Additional systems keep track of reputations

(e.g., a person's moral reputation) and information about friendships, alliances, coalitions, and benefits and burdens implicitly or explicitly accruing from social contracts (Carruthers, 2006). All the preceding systems interact to produce high-level explanations and interpretations of complex social interactions.

In our extended CLS framework, episodic memory-based learning examples stored and reactivated by the hippocampal surface system are "consumed" by these conceptual learning systems. These multiple conceptual systems separately engage in various forms of deep learning, including such things as detecting generalizations, constructing causal explanations, predicting downstream consequences, identifying cross-domain relationships and similarities, assigning social meanings, and so forth.

< Prev   CONTENTS   Source   Next >