Machine Learning for Multi-Cell Human Intent Inference

We applied modern machine learning algorithms to the study of simultaneous firings of neurons in human subjects implanted with intracranial electrodes as part of a surgical treatment for drug-resistant epilepsy. These subjects played a taxi-driver video game in which they alternated between searching for passengers, and delivering these passengers to their destination. We used support vector machines and boosting to create classification models that use multi-cell neural activity to predict whether subjects were searching for a passenger or a store. Despite methodological challenges (such as the electrodes being positioned for clinical purposes, not for our classification task), these models were able to frequently and significantly outperform a number of natural baseline (single-cell) classifiers. The SVM models are stronger than those of boosting on our problem. Furthermore, model accuracy is significantly enhanced when we consider the confidence levels output by the models (i.e. when we choose to classify only a high-confidence subset of the test data). We present compelling evidence that the learned models are integrating information from multiple cells. This work demonstrates the potential of machine learning algorithms for isolating correlates of behavior in complex neural datasets, and is one of the few such applications to human neural data for a relatively abstract cognitive state such as intent.