Wisdom of the Crowd: Using Multi-human Few-shot Learning to Improve Cross- User Generalization for Error Potentials in BCI Systems

Abstract

Brain-computer interfaces (BCI) have surfaced as a powerful and attractive modality in human-machine interaction. BCIs establish a firsthand communication channel between a human and a computer, using the human subject's neural activity signals as input. Depending on the application, BCIs use distinct brain signals, which are tailored to specific tasks. Systems designed for BCI suffer from two major problems. Firstly, the signal recorded by off-the-shelf devices is extremely noisy and suffers from a very poor signal-to-noise ratio along with low resolution and the presence of signal artifacts. This impedes achieving high accuracy in models even designed for single-user BCI systems. One method to overcome this is to utilize multiple users' observations and average their signals to get an aggregate signal. However, the second problem arises as BCI signals generalize poorly to other subjects which makes user-specific training mandatory for the model to perform well across the board or generalize well to unseen data. This places a lot of overhead on new users who want to use a BCI device in the form of calibration and training. These problems have been the subject of BCI research and various solutions have been proposed throughout the literature. In this paper, we investigate these problems and propose using multiplicity as a viable dimension along which we can find a solution. We augment our method using transfer learning (few-shot learning specifically) to address the problem of poor generalization of the BCI signal across users. We conduct experiments where multiple test subjects observe an AI agent navigate a maze in an Atari-based game environment, where the agent takes occasional incorrect actions. This generates a BCI signal called the error potential (ErrP). ErrPs have been used throughout scientific literature in cyber-physical systems to detect anomalous situations, correct robot mistakes, train machine learning algorithms, etc. However, systems designed for ErrP detection suffer from poor generalization accuracy (≈60%). In order to test the viability of using multiplicity in a single-user BCI system, we compare our performance with the baseline of the traditional single-user BCI systems and achieve a much-improved generalization accuracy <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\approx 90\%)$</tex> .