Adapting Supervised Machine Learning for Analysis of Neurobiological Signals

Machine learning algorithms are uniquely suited to decipher high-dimensional neural data, which often contains multiple overlapping signals carried by the same set of cells, leading to a low signal-to-noise ratio for individual signals. In this study, we explored the role of the retrosplenial cortex (RSC) in navigation, specifically investigating whether landmarks are encoded in this area and if they serve as an error signal enabling the RSC to adjust its internal position representation. We gathered data from hundreds of neurons while mice ran along a dark corridor with textured patches on the ground. This data was then processed using a multi-layer perceptron (MLP) to classify patch identity, patch encounter number, or all events. Furthermore, we performed a position regression analysis to test the error signal hypothesis, which failed to confirm it, and highlighted potential interpretability issues arising from the use of machine learning. Nevertheless, our findings demonstrated that even with very small and shallow networks, it is feasible to classify and regress information from raw neural data. Our research underscores the continued need for comprehensive testing and cautious interpretation in machine learning applications, especially when dealing with multifaceted and complex data such as neural signals.