Convolutional neural networks for EEG signal classification in asynchronous brain-computer interfaces
Date
2019
Authors
Forney, Elliott M., author
Anderson, Charles, advisor
Ben-Hur, Asa, committee member
Kirby, Michael, committee member
Rojas, Donald, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Brain-Computer Interfaces (BCIs) are emerging technologies that enable users to interact with computerized devices using only voluntary changes in their mental state. BCIs have a number of important applications, especially in the development of assistive technologies for people with motor impairments. Asynchronous BCIs are systems that aim to establish smooth, continuous control of devices like mouse cursors, electric wheelchairs and robotic prostheses without requiring the user to interact with time-locked external stimuli. Scalp-recorded Electroencephalography (EEG) is a noninvasive approach for measuring brain activity that shows considerable potential for use in BCIs. Inferring a user's intent from spontaneously produced EEG signals remains a challenging problem, however, and generally requires specialized machine learning and signal processing methods. Current approaches typically involve guided preprocessing and feature generation procedures used in combination with with carefully regularized, often linear, classification algorithms. The current trend in machine learning, however, is to move away from approaches that rely on feature engineering in favor of multilayer (deep) artificial neural networks that rely on few prior assumptions and are capable of automatically learning hierarchical, multiscale representations. Along these lines, we propose several variants of the Convolutional Neural Network (CNN) architecture that are specifically designed for classifying EEG signals in asynchronous BCIs. These networks perform convolutions across time with dense connectivity across channels, which allows them to capture spatiotemporal patterns while achieving time invariance. Class labels are assigned using linear readout layers with label aggregation in order to reduce susceptibility to overfitting and to allow for continuous control. We also utilize transfer learning in order to reduce overfitting and leverage patterns that are common across individuals. We show that these networks are multilayer generalizations of Time-Delay Neural Networks (TDNNs) and that the convolutional units in these networks can be interpreted as learned, multivariate, nonlinear, finite impulse-response filters. We perform a series of offline experiments using EEG data recorded during four imagined mental tasks: silently count backward from 100 by 3's, imagine making a left-handed fist, visualize a rotating cube and silently sing a favorite song. Data were collected using a portable, eight-channel EEG system from 10 participants with no impairments in a laboratory setting and four participants with motor impairments in their home environments. Experimental results demonstrate that our proposed CNNs consistently outperform baseline classifiers that utilize power-spectral densities. Transfer learning yields an additional performance improvement, but only when used in combination with multilayer networks. Our final test results achieve a mean classification accuracy of 57.86%, which is 8.57% higher than the 49.29% achieved by our baseline classifiers. In terms of information transfer rates, our proposed methods achieve a mean of 15.82 bits-per-minute while our baseline methods achieve 9.35 bits-per-minute. For two individuals, our CNNs achieve a classification accuracy of 90.00%, which is 10-20% higher than our baseline methods. A comparison with external studies suggests that these results are on par with the state-of-the-art, despite our relatively rigorous experimental design. We also perform a number of experiments that analyze the types of patterns our classifiers learn to utilize. This includes a detailed analysis of aggregate power-spectral densities, examining the layer-wise activations produced by our CNNs, extracting the frequency responses of convolutional layers using Fourier analysis and finding optimized input sequences for trained networks. These analyses highlight several ways that the patterns our methods learn to utilize are related to known patterns that occur in EEG signals while also creating new questions about some types of patterns, including high-frequency information. Examining the behavior of our CNNs also provides insights into the inner workings of these networks and demonstrates that they are, in fact, learning to form hierarchical, multiscale representations of EEG signals.
Description
Rights Access
Subject
brain-computer interfaces
electroencephalography
artificial neural networks
mental tasks
convolutional neural networks