Neural Dynamics of Category Learning and Recognition: Attention, Memory Consolidation, and Amnesia

Abstract

Abstract A theory is developed of how recognition categories can be learned in response to a temporal stream of input patterns. Interactions between an attentional subsystem and an orienting subsystem enable the network to self-stabilize its learning, without an external teacher, as the code becomes globally self-consistent. Category learning is thus determined by global contextual information in this system. The attentional subsystem learns bottom-up codes and top-down templates, or expectancies. The internal representations formed in this way stabilize themselves against recoding by matching the learned top-down templates against input patterns. This matching process detects structural pattern properties in addition to local feature matches. The top-down templates can also suppress noise in the input patterns, and can subliminally prime the network to anticipate a set of input patterns. Mismatches activate an orienting subsystem, which resets incorrect codes and drives a rapid search for new or more appropriate codes. As the learned code becomes globally self-consistent, the orienting subsystem is automatically disengaged and the memory consolidates. After the recognition categories for a set of input patterns self-stabilize, those patterns directly access their categories without any search or recoding on future recognition trials. A novel pattern exemplar can directly access an established category if it shares invariant properties with the set of familiar exemplars of that category. Several attentional and nonspecific arousal mechanisms modulate the course of search and learning. Three types of attentional mechænism—priming, gain control, and vigilance—are distinguished. Three types of nonspecific arousal are also mechanistically characterized. The nonspecific vigilance process determines how fine the learned categories will be. If vigilance increases due, for example, to a negative reinforcement, then the system automatically searches for and learns finer recognition categories. The learned top-down expectancies become more abstract as the recognition categories become broader. The learned code is a property of network interactions and the entire history of input pattern presentations. The interactions generate emergent rules such as a Weber Law Rule, a 2/3 Rule, and