Choice anticipation as gated accumulation of sensory predictions.

Abstract

Predictions are combined with sensory information when making choices. Accumulator models have conceptualized predictions as trial-by-trial updates to a baseline evidence level. These models have been successful in explaining the influence of choice history across-trials, however, they do not account for how sensory information is transformed into choice evidence. Here, we derive a gated accumulator that models the onset of evidence accumulation as a combination of delayed sensory information and a prediction of sensory timing. To test how delays interact with predictions, we designed a free-choice saccade task where participants directed eye movements to either of two targets that appeared with variable delays and asynchronies. Despite instructions not to anticipate, participants responded before target onset on some trials. We reasoned that anticipatory responses reflected a trade-off between inhibiting and facilitating the onset of evidence accumulation via a gating mechanism as target appearance became more likely. We then found that anticipatory responses were more likely following repeated choices, suggesting that the balance between anticipatory and sensory responses was driven by a prediction of sensory timing. By fitting the gated accumulator model to the data, we found that variance in within-trial fluctuations in baseline evidence best explained the joint increase of anticipatory responses and faster sensory-guided responses with longer delays. Thus, we conclude that a prediction of sensory timing is involved in balancing the costs of anticipation with lowering the amount of accumulated evidence required to trigger saccadic choice.NEW & NOTEWORTHY Evidence accumulation models are used to study how recent history impacts the processes underlying how we make choices. Biophysical evidence suggests that the accumulation of evidence is gated, however, classic accumulator models do not account for this. In this work, we show that predictions of the timing of sensory information are important in controlling how evidence accumulation is gated and that signatures of these predictions can be detected even in randomized task environments.