Latent circuit inference from heterogeneous neural responses during cognitive tasks

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Higher cortical areas exhibit complex neural responses mixing sensory, cognitive, and motor signals.
• Existing dimensionality reduction methods often overlook the circuit-level connectivity driving heterogeneous neural responses and behavior.

Purpose of the Study:

• To develop a novel dimensionality reduction approach, the latent circuit model, to infer recurrent neural connectivity underlying behavior.
• To investigate how task variables interact through low-dimensional recurrent connectivity to generate behavioral output.

Main Methods:

• Developed the latent circuit model, a dimensionality reduction technique focusing on recurrent connectivity.
• Applied latent circuit inference to recurrent neural networks trained on a context-dependent decision-making task.
• Validated model predictions through patterned connectivity perturbations and analysis of neural data.

Main Results:

• Identified a neural suppression mechanism where contextual representations inhibit irrelevant sensory responses.
• Confirmed that perturbations predicted by the latent circuit model causally affect behavior.
• Observed similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.

Conclusions:

• The latent circuit model provides insights into how neural circuit connectivity shapes behavior.
• Neural suppression of irrelevant sensory information is a key mechanism in context-dependent decision-making.
• Incorporating causal interactions is vital for understanding behaviorally relevant neural computations.