Flexible compression of sensory evidence makes human decisions more resilient to noisy neural computations

Abstract

Categorizing ambiguous stimuli requires accumulating sensory evidence over time through noisy neural computations. Theories disagree on whether this evidence is accumulated in its native sensory space or in an abstract and compressed category space that retains only decision-relevant information. Here we designed a visual categorization task in which revealing decision categories before or after a stimulus sequence determines which space can be used to accumulate evidence, without changing the optimal decision. Humans were more accurate, and their inferences less noisy, when they could accumulate evidence in category space. Concurrent magnetoencephalographic (MEG) recordings revealed flexible and accuracy-predictive representations of evidence compressed onto decision categories in these conditions. Recurrent neural networks (RNNs) trained on the same task developed the same strategy, which made their decisions more resilient to noisy computations. Together, these findings suggest that the human brain mitigates the cost of noisy inferences by focusing limited computational resources on decision-relevant information.