One-Back Perceptual Memory Task: Psychophysics, Computational Model, and EEG Correlates

To yield generalizable insights, models of cognitive function must explain behavioral and brain data across diverse experimental paradigms. In the domain of sensory-perceptual memory, however, most models are tailored to specific tasks, raising the question: are perceptual memories governed by paradigm-specific mechanisms, or can broader, unified principles be identified? A general framework in which different forms of perceptual memory arise from the interaction of multiple memory buffers with distinct dynamics has been proposed in our research group (Hachen et al. , Giana et al.). In this model, the “short-term buffer” (STB) retains recent sensory input for comparison, while the “long-term buffer” (LTB) stores stable contextual information such as categorization boundaries. Critically, the STB and LTB interact via mutual attraction. This dynamic can explain both contraction bias in working memory tasks (where short-term traces shift toward the recency-weighted longer-term prior) and repulsion effects in reference memory tasks (where judgments are biased away from the stimuli of recent-past trials). To challenge this theory with new data, in this thesis we develop a novel paradigm—the one-back memory task—which integrates delayed comparison and binary categorization. Two sequential trials can be referred to as n-1 and n. In each trial, the participant receives one vibrotactile stimulus and is tasked with judging whether that stimulus (trial n) is stronger or weaker than that of the previous trial (n-1 ). Each stimulus thus serves a dual role: as a comparison target in relation to the preceding trial and as a reference for the next trial. We hypothesize that the memory of trial n–1, held in the STB, is attracted toward the LTB across the intertrial interval between n-1 and n. In parallel, the LTB is “updated” via an attraction towards n-1 across this same interval. Although the LTB is not explicitly read out, it biases perception by acting upon the STB. Stimuli follow a deterministic Markov sequence spanning nine intensity levels. The stimuli are organized into high- and low-intensity “clouds,” with sequences of stimuli tending to occupy one cloud before jumping to the other. For reasons that will be expanded in the thesis main body, this structure allows us to investigate whether, and how, the STB and LTB jointly shape perception. To formalize these interactions, we implement a computational model in which the STB and LTB influence each other with distinct time constants (τ STB and τ LTB). Both symmetric dynamics (τ STB equal to τ LTB) and asymmetric dynamics (τ STB not constrained to be equal to τ LTB) are evaluated to determine which best accounts for behavioral findings. Finally, to identify the neuronal mechanisms underpinning these interactions, we record EEG while participants perform the task. Multivariate pattern analysis reveals dynamic activation patterns in prefrontal and posterior cortices during encoding, maintenance, and decision-making. Cluster-based analyses further uncovered context-sensitive neural signatures modulated by cloud identity, string position, and statistical transitions, indicating dynamic tracking of higher-order temporal structure. Time-frequency analyses further hints toward large-scale network dynamics spanning the somatosensory, prefrontal, and parietal cortices. These results provide new insight into how past and present sensory information are integrated by interacting memory systems to guide perceptual decisions.