Neural network dynamics underlying context-dependent perceptual decision making

In perceptual judgments, sensory information is typically combined with contextual information which can determine the prior statistics. In a previously developed 2AFC auditory categorization task in which the stimulus sequence included two different statistical contexts, rats can capitalize on the context information but only in after-correct trials; in after-error trials, they consistently deviate from optimal behavior by waiving the accumulated context evidence (Hermoso-Mendizabal et al. 2020). Recently, it has been shown that Recurrent Neural Networks (RNNs) pre-trained in a naturalistic task presenting more than two possible choices reproduce the animal’s suboptimal behavior in the 2AFC, which has been termed after-error reset strategy (Molano-Mazón et al. 2021).
To what extent the neuronal mechanisms underlying the representation of sequential trial statistics and the reset strategy are shared between these pre-trained RNNs and animals, remains an open question.
To address this question, we performed a population analysis of pre-trained RNNs and found that they formed an accurate representation of the statistical context independent of the previous outcome. After error trials, the reset was implemented nonetheless by dynamically decoupling the network’s decisions from the predictions based on the context evidence. We conducted the same analysis using neural recordings from the dorsomedial striatum (DMS) and frontal orienting field (FOF) of rats performing the task. The statistical context was encoded in the DMS, but at output-level, this encoding did not affect the rats’ choices in after-error trials. However, unlike in the RNNs where the previous choices were always well encoded, there was a loss of information in the rats’ neural activity concerning the previous incorrect choices. The FOF in contrast showed only a weak encoding of context. Our comparative population analysis in RNNs and rats' striatal activity provides an integrated insight into the neural computational mechanisms that support flexible behavior in context-dependent perceptual tasks.