A unified model for recognition and prediction using a compressed internal timeline

Abstract

It has long been understood that there is a deep connection between time and memory. From episodic memory in humans to conditioning tasks in animals, temporal relationships play a crucial role in memory performance. While recognition memory is a subset of episodic memory, most recognition memory models disregard information about time and assume that memory is a composite store with a noisy record of items and their associations. Another class of models posits that memory depends on temporal representations in which ‘what’ and ‘when’ information is stored conjointly. Using three experiments, I found evidence for serially accessing memory (scanning) in both short-term and long-term memory and in predicting the future. These findings support the hypothesis that memories are stored in temporal representations. In Experiment 1, I hypothesized that scanning in a judgment-of-recency task is due to a compressed temporal representation. In 107 healthy young adults, response times depended only on the lag to the target and varied sub-linearly with lag. This result was consistent with the hypothesis. In Experiment 2, the hypothesis was that memory search on a long-term recognition task is driven by serially scanning a compressed representation. In a continuous recognition paradigm with 88 healthy young adults across three studies, the time at which information starts becoming accessible varied as a function of the logarithm of the lag. This result suggests that information in long-term memory is stored in a compressed representation that can be accessed using a serial backward scan. In Experiment 3, I tested the hypothesis that our ability to access what is going to happen a few seconds in the future is similar to our ability to access the immediate past. Sixty healthy young adults performed a relative order judgment task for future events. The response times in this novel judgment-of-imminence task showed that a search through prospective memory representation was serial and closely paralleled the serial search observed in the judgment-of-recency task (Experiment 1). Together, these results suggest that it is possible to generate a temporally ordered representation that can be scanned to access the past and the future.