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

Abstract

Decoding visual information based on machine learning has the potential to reveal the neural mechanisms of visual information processing in the brain. Although most highly precise decoding methods require highly invasive electrodes, these electrodes make long-term use difficult because they cause brain damage. To address this issue, we recorded local field potentials (LFPs) from the mouse cortex using custom-made low-invasive electrodes (ECoG with six recording sites) and verified how convolutional neural network (CNN) can decode visual stimulus positions from the LFP signals recorded by these electrodes. We found that the discrimination accuracies based on the cross-validation (39-70%) were significantly higher than the chance level (25%) for all animals. This confirmed that CNNs can extract the neural features of visual stimulus positions, even from low-spatial-resolution LFPs. Furthermore, we showed that visual positional information was primarily encoded during specific time periods of LFPs by using surrogate datasets. Our findings suggest that CNN enables low-invasive electrodes to analyze the visual information processing system and provide tools to understand the neural mechanisms in the brain.