Latent circuit inference from heterogeneous neural responses during cognitive tasks

Abstract

Higher cortical areas carry a wide range of sensory, cognitive, and motor signals supporting complex goal-directed behavior. These signals are mixed in heterogeneous responses of single neurons tuned to multiple task variables. Dimensionality reduction methods used to analyze neural responses rely merely on correlations, leaving unknown how heterogeneous neural activity arises from connectivity to drive behavior. Here we present a framework for inferring a low-dimensional connectivity structure—the latent circuit—from high-dimensional neural response data. The latent circuit captures mechanistic interactions between task variables and their mixed representations in single neurons. We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses. We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit structure. Our approach can reveal interpretable and causally testable circuit mechanisms from heterogeneous neural responses during cognitive tasks.

Relating local connectivity and global dynamics in recurrent excitatory-inhibitory networks

“…We therefore opted for an approximation based on truncated eigen-decomposition. When studying input-driven and transient dynamics, different methods for low-rank approximation may be more appropriate, and are a topic of active research [57–61]…”

“…When studying input-driven and transient dynamics, different methods for low-rank approximation may be more appropriate, and are a topic of active research [57][58][59][60][61] To perform the eigen-decomposition of excitatory-inhibitory connectivity matrices, we leveraged the fact that they can be expressed as a sum of a block-like deterministic low-rank matrix and a full-rank random matrix with zero-mean [40]. The eigen-spectrum of such matrices in general consists of a continuously-distributed bulk and discrete outliers [44,45].…”

“…When studying input-driven and transient dynamics, different methods for low-rank approximation may be more appropriate, and are a topic of active research [57][58][59][60][61] To perform the eigen-decomposition of excitatory-inhibitory connectivity matrices, August 25, 2022 20/47…”

Opening Questions in Visual Working Memory

“…Methods borrowed from other disciplines help analyse and compare structures, dynamics, and trajectories of functional networks. Computational models have increased their sophistication and physiological validity (e.g., Langdon & Engel, 2022).…”

Beyond accuracy: generalization properties of bio-plausible temporal credit assignment rules

“…The error gradient, if available, tells us how each parameter should be adjusted in order to have the steepest local descent. For training RNNs, which are widely used as a model for neural circuits [44][45][46][47][48][49][50][51][52][53][54][55][56][57][58][59][60][61][62][63], standard algorithms that follow this gradientreal time recurrent learning (RTRL) and BPTT -are not bio-plausible and have overwhelming memory storage demands [3,64]. However, learning rules that only approximate the true gradient can sometimes be as effective as those that follow the gradient exactly [10,65].…”