Exploiting correlations across trials and behavioral sessions to improve neural decoding.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning in Biology

Background:

• Traditional neural decoders analyze single trials, ignoring crucial cross-session and cross-trial neural activity patterns.
• Animal behavior is influenced by prior experiences and exhibits consistent neural patterns during repeated tasks.

Purpose of the Study:

• To develop novel computational models that integrate neural activity across multiple sessions and trials.
• To improve the decoding of behavior from neural data by accounting for temporal dependencies.

Main Methods:

• Introduced two complementary multi-session models: reduced-rank regression and state-space models.
• Applied models to a large dataset (433 sessions, 270 brain regions) from the International Brain Laboratory (IBL) mouse Neuropixels dataset.

Main Results:

• The proposed decoders significantly outperformed traditional methods in decoding four distinct behaviors.
• Model performance generalized across different datasets, species, and experimental tasks.
• Achieved efficient and interpretable neural representations, identifying task-related single-neuron contributions and brain-wide activation timescales.

Conclusions:

• Multi-session modeling approaches are superior to single-session methods for understanding neural correlates of behavior.
• These models provide interpretable, low-dimensional neural representations valuable for neuroscience research.
• The developed methods offer a powerful, efficient alternative to deep learning for neural decoding.