Poster Presentation

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Optimally Structured Mixed Selectivity. A Normative Theory of Control in Neural Circuits

Will Dorrell¹; ¹Harvard University

Presenter: Will Dorrell

Neuroscientists and psychologists have long-puzzled over representations of task-relevant variables in PFC, seat of the mind’s executive control (Miller & Cohen, 2001), working memory (Funahashi et al., 1993), and schematic knowledge (Baldassano et al., 2018). We can summarise this work by contrasting the two proposed representations: modular vs. conjunctive. Modular codes neurally separate components of the task, a boon for compositional generalisation, but not for multi-tasking nor flexible downstream decoding (Badre, 2025; Feng et al., 2014; Musslick & Cohen, 2019, 2021). Answering both these concerns, conjunctive codes, in which single neurons are non-linearly mixed-selective to multiple task factors, have become popular (Rigotti et al., 2010, 2013; Tye et al., 2024), and align with PFC’s complex neural responses (El-Gaby et al., 2024; Warden & Miller, 2010; Xie et al., 2022). However, while PFC does exhibit high degrees of mixed-selectivity, its representations are also highly structured (Baram et al., 2021; Chiang & Wallis, 2018; El-Gaby et al., 2024; Samborska et al., 2022; Tian et al., 2024). Here, instead, we draw insight from the successes of circuit models for path-integration, like the fly head direction (Kim et al., 2017; Turner-Evans et al., 2017) or mammalian grid cell system (Rowland et al., 2016). These circuits carefully choreograph the connectivity between modular (e.g. grid cells) and conjunctive (e.g. conjunctive grid cells) neurons to implement meaningful computations. In these circuits, conjunctive neurons form the substrate of internal control, updating the representation of position in response to movements. We argue that this mantra - conjunction for control - is much broader. We show empirically that recurrent neural networks (RNNs) trained to track Finite-State Machines (FSMs) display similar tuning; we theoretically justify this with novel link between regularised RNNs and convex neural networks; and use this link to reinterpret recently-reported ’temporary variable’ subspaces. In sum, we argue for reinterpreting conjunctive coding as a symptom of internal neural control, rather than flexible downstream decoding. In so doing, we provide another tool in the ongoing quest to decode neural computations, in PFC and elsewhere, from available measurements.

Topic Area: Methods, Tools, Theory & Neural Coding