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

Area of Science:

• Neuroscience
• Machine Learning
• Computational Neuroscience

Background:

• Temporal distributional shifts in data can hinder accurate early classification of time-series data.
• Recurrent Neural Networks (RNNs) are widely used for sequence data but struggle with temporal shifts.
• Early detection of behavior from neural activity is crucial for timely interventions.

Purpose of the Study:

• To introduce Time-Varying RNNs (TV-RNNs) for robust early classification of time-series neural data.
• To enhance RNN memory and temporal feature utilization for improved classification accuracy.
• To apply TV-RNNs to neural data for early behavioral classification in motor tasks.

Main Methods:

• Developed and implemented Time-Varying RNNs (TV-RNNs) with time-varying weights.
• Applied TV-RNNs to diverse neural activity datasets from mice and humans during motor tasks.
• Utilized SHapley Additive exPlanation (SHAP) values to analyze brain region contributions.

Main Results:

• TV-RNNs achieved accurate classification earlier in the time-series compared to standard RNNs.
• The models demonstrated robustness in classifying neural data despite temporal distributional shifts.
• SHAP analysis identified somatosensory and premotor regions as critical for behavioral classification.

Conclusions:

• TV-RNNs offer a significant advancement for early sequential classification of neural data.
• The proposed method enhances the ability to decode behavior from neural activity in real-time.
• Understanding brain region contributions aids in developing targeted neural stimulation strategies.