Mengting Fang¹ (), Jiang Mao¹, Tobias H. Donner², Alan A. Stocker¹; ¹University of Pennsylvania, ²University Medical Center Hamburg-Eppendorf

Temporal integration of sensory information is an important aspect in many perceptual decision tasks. Our prior study found that a temporal break in the middle of an otherwise regular sequence of stimulus samples resulted in an increased weighting of the first post-break sample when performing a subsequent perceptual judgment. To further probe the origin of this phenomenon, we conducted a visual estimation experiment with varying durations and positions of the break in the sample sequence. Subjects estimated the angular position of an unknown generative mean based on 8 normally distributed stimulus samples of that mean (fixed variance). The samples were presented in rapid sequence. Subjects reported their estimates using a joystick and were given immediate feedback by showing the true generative mean. Subjects performed the task under three conditions. First, stimulus samples were presented without interruption. Second, in the "break" condition, sample sequences were interrupted after the 4th sample for varying break durations. Finally, in the "break" condition with different break positions, interruptions occurred randomly after either the 2nd, 4th, or 6th sample (1.75s duration). We calculated the relative contribution of each sample to the final estimate. Results showed that the weight of the first sample after the break monotonically increased with break duration, implying a continuous underlying process rather than a threshold mechanism. Furthermore, over-weighting persisted for other break positions in the sequence. We developed a normative model that assumes an active management of the trade-off between estimation performance on one hand and stimulus encoding and memory maintenance efforts on the other. This model effectively captures the temporal dynamics of the integration process from all our experimental data in conditions with or without breaks, which other models can not (e.g., bump attractor networks). Our results suggest that perceptual evidence accumulation is an actively controlled, resource-rational process.

Acknowledgements: This work was supported by the NSF CRCNS grant IIS-1912232 to A.A.S and Federal Ministry of Education and Research (BMBF) Project-nr: 01GQ1907 to T.H.D.