Systems-level Models of Working Memory: The Role of Reward-based Learning

Ahmed Moustafa

Institute of Cognitive Science

University of Louisiana at Lafayette

 

Abstract

Different brain areas subserve performance of working memory (WM) tasks, including the basal ganglia and prefrontal cortex (PFC). A class of WM tasks, known as delayed-response tasks, has been used to assess the relevance of the basal ganglia or PFC to WM in nonhuman animals (Collins et al., 2000; Goldman-Rakic, 1995a, 1995b; Levy et al., 1997; Sawaguchi & Iba, 2001), as well as patients with basal ganglia disorders (Vermersch et al., 1999; Lewis et al., 2003). We propose a neural model that simulates performance of delayed-response tasks. An architecture, known as the Actor-Critic architecture (Houk et al., 1995), has been used to model basal ganglia-based motor and/or cognitive functions (Berns & Sejnowski, 1996; Suri & Schultz, 1998, 1999; Suri et al., 2001). We incorporate this architecture to model performance in delayed-response tasks. The model is trained using a reward-based learning algorithm known as the temporal difference algorithm. The model assumes that PFC subserves maintenance of information in WM, while the basal ganglia subserve selection of motor- and cognitive-related information (Prescott et al., 2003; Redgrave et al., 1999). In this talk, I will discuss experimental and modeling studies (including the proposed model) related to performing delayed-response tasks.