Deep Learning for Mind Reading: Using Neural Networks to Forecast Neural Signals

Abstract

Brain-computer interfaces have seen unprecedented advances during the past decade. A particularly interesting area of research is related to speech neuroprostheses: devices that can translate thoughts directly into speech or text. This work contributes to the development of speech neuroprostheses by attempting to forecast brain signals recorded using electrocorticography (ECoG). The applications of this work include speech forecasting, the modeling of speech producing areas in the brain, and providing context to models used for brain-to-speech decoding. We use different neural network models and find that ECoG forecasting is possible with mixed results. While neural network models can predict a trend associated with the data, modeling the specific amplitudes proved more difficult. We finish by suggesting a few models that could be used to improve speech neuroprosthesis research.