Inferring Latent Neural Circuits from Heterogeneous Neural Responses During Cognitive Tasks

Academic Background

In cognitive tasks, higher cortical areas of the brain (such as the prefrontal cortex, PFC) are responsible for integrating a variety of sensory, cognitive, and motor signals. However, the responses of individual neurons often exhibit complex heterogeneity, meaning they respond simultaneously to multiple task variables. This heterogeneity makes it difficult for researchers to directly infer the neural circuit mechanisms driving behavior from neural activity. Traditional dimensionality reduction methods rely on correlations between neural activity and task variables but fail to reveal the neural circuit connections behind these heterogeneous responses.

To address this issue, Christopher Langdon and Tatiana A. Engel developed a new dimensionality reduction method—the Latent Circuit Model. This model simulates interactions among task variables through low-dimensional recurrent connectivity and generates behavioral outputs. Through this model, researchers can infer low-dimensional neural circuit mechanisms from high-dimensional neural response data, thereby uncovering computational processes in cognitive tasks.

Source of the Paper

This paper was co-authored by Christopher Langdon and Tatiana A. Engel, who are affiliated with the Princeton Neuroscience Institute at Princeton University and Cold Spring Harbor Laboratory, respectively. The paper was published in Nature Neuroscience in 2025, titled “Latent circuit inference from heterogeneous neural responses during cognitive tasks.”

Research Workflow and Results

1. Development of the Latent Circuit Model

The authors first developed the latent circuit model, which simulates interactions among task variables through low-dimensional recurrent connectivity. The core idea of the model is to map high-dimensional neural response data into a low-dimensional latent space and generate behavioral outputs through a low-dimensional neural network. Specifically, the dynamic equation of the model is as follows: $$ \dot{x} = -x + f(W{rec}x + W{in}u) $$ Here, (x) represents the latent variable, (W{rec}) is the recurrent connectivity matrix, (W{in}) is the input connectivity matrix, (u) is the task input, and (f) is the activation function.

2. Application of the Model to RNNs

To validate the effectiveness of the latent circuit model, the authors applied it to a trained Recurrent Neural Network (RNN) that was trained to perform a context-dependent decision-making task. In this task, the RNN needed to selectively process sensory inputs based on contextual cues. Using the latent circuit model, the authors discovered an inhibitory mechanism where contextual representations suppress irrelevant sensory responses. This mechanism aligns with hypotheses from previously hand-crafted neural circuit models.

3. Validation of the Model

To validate the mechanisms inferred by the latent circuit model, the authors performed patterned perturbations on the RNN’s connectivity and observed their effects on behavior. The results showed that the effects of these perturbations on behavior were consistent with the predictions of the latent circuit model, further confirming the inferred inhibitory mechanism. Additionally, the authors also discovered a similar inhibitory mechanism in prefrontal cortex neural recordings from monkeys performing the same task.

4. Comparison with Regression Models

The authors also compared the latent circuit model with traditional regression models. Regression models analyze neural responses by finding low-dimensional projections most correlated with task variables but do not consider interactions among task variables. The results showed that regression models failed to reveal the inhibitory mechanism discovered by the latent circuit model, indicating that ignoring interactions among task variables may lead to misunderstandings of neural computation.

Conclusions and Significance

Through the development of the latent circuit model, this study successfully inferred low-dimensional neural circuit mechanisms from high-dimensional neural response data. This model not only explains the origins of heterogeneous neural responses but also reveals computational processes in cognitive tasks. Furthermore, the study demonstrates that low-dimensional neural circuit mechanisms in high-dimensional networks can still drive behavior. This discovery provides a new perspective for understanding cognitive functions in the brain and offers a powerful tool for future neural circuit research.

Research Highlights

1. Novel Model Development: The latent circuit model is the first to combine low-dimensional neural circuit mechanisms with high-dimensional neural response data, bridging the gap between traditional dimensionality reduction methods and neural circuit mechanisms.
2. Discovery of Inhibitory Mechanism: Through model inference, the authors discovered an inhibitory mechanism where contextual representations suppress irrelevant sensory responses. This mechanism was validated in both RNNs and the prefrontal cortex of monkeys.
3. Comparison with Traditional Models: By comparing with regression models, the study emphasizes the importance of interactions among task variables, providing new directions for future neural computation research.

Other Valuable Information

The study also demonstrates how to validate inferred neural circuit mechanisms through perturbation experiments, offering actionable methods for future neural circuit research. Additionally, the results suggest that low-dimensional mechanisms in high-dimensional networks may be similar to classical small-scale neural circuit models, providing new insights into understanding complex neural computations.