Convolutional neural networks decode visual stimulus positions from local field potentials on the mouse cortex

Area of Science:

• Neuroscience
• Machine Learning
• Computational Neuroscience

Background:

• Decoding neural activity is crucial for understanding brain function.
• Highly precise decoding methods often rely on invasive electrodes, limiting long-term studies.
• There is a need for non-invasive or low-invasive techniques for brain activity analysis.

Purpose of the Study:

• To investigate the efficacy of convolutional neural networks (CNNs) in decoding visual stimulus positions.
• To assess the feasibility of using low-invasive electrocorticography (ECoG) with CNNs for brain activity analysis.
• To explore the temporal dynamics of visual positional information encoding in the cortex.

Main Methods:

• Recorded local field potentials (LFPs) from the mouse cortex using custom low-invasive ECoG electrodes (six recording sites).
• Applied CNNs to decode visual stimulus positions from the recorded LFP signals.
• Utilized cross-validation for accuracy assessment and surrogate datasets to analyze temporal encoding.

Main Results:

• CNNs achieved significantly higher discrimination accuracies (39-70%) than chance level (25%) for decoding visual stimulus positions.
• CNNs successfully extracted neural features of visual positions even from low-spatial-resolution LFPs.
• Visual positional information was found to be encoded during specific time periods within the LFP signals.

Conclusions:

• CNNs enable effective decoding of visual information using low-invasive ECoG recordings.
• This approach provides a valuable tool for studying neural mechanisms of visual processing with reduced invasiveness.
• Findings pave the way for advanced, long-term investigations into brain function.