Behavioral Classification of Sequential Neural Activity Using Time Varying Recurrent Neural Networks.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Temporal distributional shifts in time-series data challenge early classification.
• Accurate early decoding of neural activity is crucial for timely interventions, like corrective neural stimulation.
• Standard recurrent neural networks (RNNs) struggle with temporal shifts and lack robust long-term memory.

Purpose of the Study:

• To introduce a novel RNN architecture, Time-varying RNNs, designed to handle temporal distributional shifts.
• To enhance RNNs' ability to utilize all temporal features and improve memory for sequence data.
• To achieve earlier and more robust classification of time-series data, specifically neural activity.

Main Methods:

• Developed Time-varying recurrent neural networks (TV-RNNs) with time-varying weights.
• Applied TV-RNNs to classify spatially distributed neural activity from motor tasks in mice and humans.
• Utilized SHapley Additive exPlanation (SHAP) values to analyze brain region contributions to classification.

Main Results:

• TV-RNNs achieved accurate classification earlier in the sequence compared to standard RNNs.
• TV-RNNs demonstrated robust classification despite temporal distributional shifts.
• Early detection of self-initiated lever-pull behavior was improved by up to 3 seconds (6 seconds before onset).
• SHAP analysis identified somatosensory and premotor regions as critical for behavioral classification.

Conclusions:

• Time-varying RNNs offer a significant advancement for early sequential classification of neural data.
• These models provide more stable gradient dynamics and enhanced memory compared to standard RNNs.
• The findings highlight the importance of somatosensory and premotor cortex in motor behavior classification.