How Agents Learn to Manage the Speed-Accuracy Tradeoff

Abstract

One way to understand how the brain generates behavior is to measure behavior in controlled conditions and infer the underlying neural mechanisms. This is the method employed for the study of two salient types of behavior described in this thesis, learning and decision-making, and invariant visual object recognition. In a study of perceptual learning and decision-making carried out in rats and complemented by modeling work, I developed a theory for learning that shows that a strategy focused on maximizing present rewards is detrimental to learning, and consequently to maximizing total reward over time. Prioritizing learning and adopting a strategy of slow initial responses led to higher total rewards, but at the cost of initial rewards rates. Rats managed this tradeoff between learning speed and instantaneous reward rate and harvested a near-optimal amount of reward. Further, learning speed scaled with initial response time both in enforced and voluntary cases. Rats only invested in learning when learning was possible, opting to maximize present rewards when learning was not possible. These results suggest that rats exhibit cognitive control of learning and may provide a principled reward-based account for the Law of Practice, the observation that reaction times decrease and accuracy increases with practice, maximizing total reward accrued over engagement with a task. A series of studies of invariant visual object recognition elaborate on rats’ learning abilities. One provides a protocol for an accelerated behavioral training regime that reduced training time severalfold over published results, while another explores the boundaries of object recognition behavior through the presentation of numerous stimulus transformations. Rats were capable of learning new stimuli without forgetting old ones, supporting the hypothesis that they learn new stimuli in relation to familiar stimuli. Results from lesion studies showed the importance of visual cortical areas for invariant object recognition in rats. Finally, I present a method for the quantitative characterization of brain lesions based on micro-computed tomography (micro-CT). Together, these studies contribute to our understanding of the complex behavioral abilities in rodents, suggesting they are capable of sophisticated strategies and are sensitive to learning contexts.