FFF CONFERENCE CTF07

Lawrence W. Barsalou - Grounding Symbolic Operations in the Brain’s Modal Systems

A central theme of modern cognitive science is that symbolic interpretation underlies human intelligence. The human brain does not simply register images, as in cameras or other recording devices. Instead symbolic interpretation of experiential content is essential for intelligent activity. What cognitive operations underlie symbolic interpretation? Across decades of analysis, a consistent set of symbolic operations has arisen repeatedly in logic, linguistics, and knowledge engineering: binding types to tokens; binding arguments to values; drawing inductive inferences from category knowledge; predicating properties and relations of individuals; combining symbols to form com-plex symbolic expressions; representing abstract concepts that interpret meta-cognitive states. It is difficult to imagine performing intelligent computation without these op-erations. For this reason, many theorists have argued that symbolic operations are cen-tral, not only to artificial intelligence, but to human intelligence.
Although classic symbolic approaches are still widely accepted as accounts of hu-man intelligence, and also as the engine for artificial intelligence, they have come in-creasingly under attack for two reasons. First, classic symbolic approaches have been widely criticized for not being sufficiently statistical. As a result, neural net approaches have developed to remedy this deficiency. Second, classic symbolic approaches have been criticized for not being grounded in perception, action, and introspection. As a result, researchers have argued that higher-order cognition is grounded in the brain’s modal systems. As statistical and grounded approaches have become increas-ingly embraced, the tendency to throw the baby out with the bath water has often been embraced as well. Some researchers have suggested that classic symbolic operations are irrelevant to higher cognition, especially researchers associated with neural nets and dynamical systems.
An alternative account of symbolic operations has arisen in grounded theories. Not only does this account have psychological and neural plausibility, it suggests a new approach to image analysis. Essentially, this approach develops symbols whose con-tent is extracted from experiential images, not necessarily conscious, distributed over space and time. As a result, symbols can be bound to regions of experience, thereby establishing type-token mappings. Inferences drawn from category knowledge also take the form of images, such that they can be mapped back into perception (the im-ages that deliver inferences take the form of simulations in the brain’s modal systems and are typically sketchy and biased, not complete and perfectly veridical). Analysis of a perceived individual proceeds by processing its space-time regions and assessing whether perceptually grounded properties and relations can be predicated of them. Symbol combination involves the manipulation and integration of image components to construct structured images that, in effect, implement complex symbolic proposi-tions. Abstract concepts result from situated introspection, namely, the process of per-ceiving internal mental and bodily states in the context of external situations, and de-veloping image-based representations of them for later use in reasoning. Behavioral and neural evidence supporting this framework will be presented.