Computational Models Suggest that Human Memory Judgments Exhibit Interference due to the Use of Distributed Representations

Abstract

Episodic memory is a core function that allows us to remember the events of our lives. Given that many events in our life contain overlapping elements (e.g., similar people and places), it is critical to understand how well we can remember the specific events of our lives vs. how susceptible we are to interference between similar memories. Decades of research have aimed to uncover the mechanisms of memory using a combination of behavioral tasks, neuroscientific measures, and computational modeling. Several prominent theories converged on the notion that the brain employs distinct neural mechanisms to support different aspects of episodic memory. Many theories suggest that the hippocampus plays a greater role in processes such as recollection, associative memory, and memory for specific details, while the neocortex plays a more prominent role in domain-specific memory (e.g., perirhinal cortex and object-based memory). Specifically, many studies have supported the role of specific subregions of the hippocampus in the computational mechanism of pattern separation, thus potentially allowing us to store and retrieve unique memories for similar experiences of our lives. However, a critical unaddressed question remains: What is the extent to which we make memory-based decisions using the purported pattern-separated representations in the hippocampus vs. other neural mechanisms that employ more distributed representations? We propose that studying human memory performance on tasks with targets and similar lures provides a critical testbed for comparing the competing predictions the role of hippocampal pattern separation vs. more distributed representations in supporting human episodic memory. We generated predictions from competing computational models of performance on memory tests with targets and similar lures and then we tested these predictions in a large sample of human participants (N=145). We found that the comparison between simulated neural responses in an object-processing region of the brain (area IT) and human memory performance exhibited a linear relationship, thus better supporting the predictions of distributed memory models than models of hippocampal pattern separation. Likewise, we observed strong effects of test format on performance as well as clear and consistent relationships between test formats, and these results were also better accounted for by the distributed memory models than the proposed pattern-separated representations of the hippocampus. Altogether, our results provide an important challenge to prominent theories of human memory and provide an important alternative mechanism for explaining human memory performance; therefore, we discuss the implications of our results for reinterpreting previous behavioral, neuroscientific, and computational modeling research and we propose avenues for future research.

Prolonged development of forced-choice recognition when targets are paired with non-corresponding lures

“…However, because MINERVA 2 assumes that item memories are stored as distributed representations, global matching models are consistent with computational models that propose that the brain represents information via distributed patterns of activity across neurons (cf. Norman and O'Reilly, 2003;Huffman and Guan, 2023). Therefore, the predictions that age-related differences in the learning rate and encoding variability account for FORCED-CHOICE RECOGNITION 30 developmental differences in mnemonic discrimination during childhood could be explored using psychophysiological (e.g., eye-tracking; Hannula et al, 2010) and neuroimaging measures, including functional magnetic resonance imaging (fMRI; Norman et al, 2006) and event-related potentials (ERPs; Rugg & Curran, 2007).…”