Using computer models to model how our memories help us learn new things from our experiences
Memory retains information that is currently not available in the environment, but memory is not a purpose in itself for a living being. The stored information confer an evolutionary benefit on the organism. The function of memory in general, and episodic memory (EM) in particular, has received far less attention than the properties the memory processes and their underlying neural mechanisms. In this subproject, we focus on directive functions, which refers to cases where the retrieved memory is used to solve a practical problem. In the first funding period, we adopted reinforcement learning for modeling the functional role of EM in spatial learning. We considered three learning paradigms: 1. EM is retrieved to learn from single experiences (one-shot learning), 2. EM is replayed to facilitate learning of statistical regularities (replay learning), and 3. learning occurs online as experiences occur, but does not have access to past experiences (online learning). We found that whether an agent is able to solve a task by relying on the three learning paradigms depends differently on the number of learning trials and the complexity of the task. EM can confer a major benefit on spatial learning, but does not always do so, and its effect differs for the two modes of accessing episodic information. One-shot learning is initially faster than replay learning, but the latter reaches a better asymptotic performance. In the second funding period, we will build on this approach, incorporating more cognitive processes associated with EM and extending it to study memory encoding and consolidation. First, by separating image processing from the reinforcement learning component, we will study how the representational format of the inputs influences learning. In addition, we will integrate a model of semantic completion to study the functional role of the generativity of EM. Second, we will test our hypothesis that attentional selection of the most informative parts of the input speeds up learning and makes behavior more robust. We will also study how attentional selection of different representational formats might make learning more flexible. Third, we will use replay prioritization to generate a variety of replay types reported in the literature to study how replay might be optimized for different task demands. Fourth, we will investigate the functional implications of EM updating. We expect that, in a changing environment, it is beneficial to increase the impact of recent memories relative to more remote memories, which gives rise to memory updating. In summary, our research will advance our understanding of the function of EM, i.e., how it drives behavior. This is an important step towards elucidating the nature of EM.