A value-based explanation for lapses in perceptual decisions

During perceptual decisions, even well-trained subjects can have a constant rate of errors independent of evidence strength, assumed to arise from inattention or motor errors. These are referred to as “lapses”, and their proper treatment is crucial for accurate estimation of perceptual parameters, however the factors influencing them remain poorly understood. Here, we propose uncertainty-guided exploration as an underlying cause for lapses. We demonstrate that perceptual uncertainty modulates the probability of lapses both within and across modalities on a multisensory discrimination task in rats. These effects cannot be accounted for by inattention or motor error, however they are concisely explained by a normative model of uncertainty-guided exploration. Further, we show that increasing the reward for one decision over the other shifts the lapse probability towards that decision in uncertain conditions, while leaving "sure-bet" decisions unchanged, as predicted by the model. Finally, we demonstrate that muscimol inactivations of secondary motor cortex and posterior striatum affect lapses across modalities. The inactivations are captured by subtractive changes to action value in the model, and do not affect "sure-bet" decisions. Together, our results suggest a value-based account for lapses, and that far from being a nuisance, lapses are informative about individual animals’ exploration-exploitation tradeoff.