A new theoretical framework jointly explains behavioral and neural variability across subjects performing flexible decision-making

Abstract

The ability to flexibly select and accumulate relevant information to form decisions, while ignoring irrelevant information, is a fundamental component of higher cognition. Yet its neural mechanisms remain unclear. Here we demonstrate that, under assumptions supported by both monkey and rat data, the space of possible network mechanisms to implement this ability is spanned by the combination of three different components, each with specific behavioral and anatomical implications. We further show that existing electrophysiological and modeling data are compatible with the full variety of possible combinations of these components, suggesting that different individuals could use different component combinations. To study variations across subjects, we developed a rat task requiring context-dependent evidence accumulation, and trained many subjects on it. Our task delivers sensory evidence through pulses that have random but precisely known timing, providing high statistical power to characterize each individual's neural and behavioral responses. Consistent with theoretical predictions, neural and behavioral analysis revealed remarkable heterogeneity across rats, despite uniformly good task performance. The theory further predicts a specific link between behavioral and neural signatures, which was robustly supported in the data. Our results provide a new experimentally-supported theoretical framework to analyze biological and artificial systems performing flexible decision-making tasks, and open the door to the study of individual variability in neural computations underlying higher cognition.

Orthogonal neural encoding of targets and distractors supports multivariate cognitive control

“…In contrast to these stark task-related differences in coherence encoding, we found that neural encoding of target response information (response-signed color coherence in the Attend-Color task and response-signed motion coherence in the Attend-Motion task) was preserved across tasks, including within dACC, SPL, and IPS (Figure 5B). Consistent with previous experiments examining context-dependent decision-making (Aoi et al, 2020;Flesch et al, 2022;Kayser et al, 2010b;Mante et al, 2013;Pagan et al, 2022;Takagi et al, 2021), we found stronger target response encoding relative to distractor response encoding, in our case in the response-encoding SPL (Attend-Color: t(28) = 4.26, one-tailed p = 0.0001; Attend-Motion: t(28) = 2.37, one-tailed p = 0.0124). We also found that target response encoding during Attend-Motion was aligned with Attend-Color, both for motion response encoding ('stimulus axis'; SPL: one-tailed p = .0236, IPS: one-tailed p = .0166) and target response encoding ('decision axis'; SPL: one-tailed p = 1.29 × 10 -6 ; IPS: one-tailed p = .0005), again in agreement with these previous experiments.…”

“…Previous work has proposed that independent neural representations play an important role in cognitive control, but have largely examined how these representations minimize interference between tasks. When tasks are in conflict, the brain uses orthogonal task representations to minimize cross-talk (Flesch et al, 2022;Kaufman et al, 2014;Mante et al, 2013;Minxha et al, 2020;Pagan et al, 2022;Panichello and Buschman, 2021;Salinas, 2004), consistent with the optimal strategy in artificial neural networks (Flesch et al, 2022;Mante et al, 2013;Musslick et al, 2020). A compelling possibility is that the cognitive control system uses a similar representational format to coordinate multiple control signals within a task as well (Ebitz et al, 2020;Libby and Buschman, 2021;Rust and Cohen, 2022).…”

Inferring context-dependent computations through linear approximations of prefrontal cortex dynamics

“…2b). The mechanisms we identified as plausible explanations of the PFC responses share key features with mechanisms of context-dependent integration that were recently described in rats 48 . Notably, that study demonstrated the advantage of pulsatile inputs in distinguishing between different mechanisms of input selection and integration.…”

Dynamics of random recurrent networks with correlated low-rank structure

“…This finding is vital for the use of RNNs as hypothesis generators [3,63,69], where it is often implicitly assumed that training results in universal solutions [33] (even though biases in the distribution of solutions have been discussed [66]). Here, we show that a specific control knob allows to move between qualitatively different solutions of the same task, thereby expanding the control over the hypothesis space [43,68]. Note in particular that the default initialization in standard learning frameworks has large output weights, which results in oblique dynamics (or unstable solutions if training without noise, see Methods, Section 4.6).…”

“…Such a decoupling for oblique, but not aligned, dynamics leads to a prediction regarding the universality of solutions [33,43,68]. For aligned dynamics, the coupling implies that the internal dynamics are strongly constrained by the task.…”