Recognizing Thoughts from Bioelectric Patterns? A Brain-Computer Interface with Deep Learning¶

J. Laiho — M.Sc. thesis, Aalto University, August 2020 (with Accenture Liquid Studio, NL)

A brain-computer interface reads electrical activity from the scalp and turns it into control signals, letting someone steer a wheelchair or a cursor by imagining a movement. The obstacle to deployment is not accuracy in the lab but calibration time. Electroencephalography (EEG) signals vary so much between people that each new user typically spends weeks training the system on their own brain. This thesis asks whether a deep network can cut that cost, and in particular whether it can work on a person it has never seen.

The learning problem¶

Data points. One data point is a single motor-imagery trial: a 2-second window, experimentally selected from the 2–4 s segment of each 8-second trial, recorded across 22 EEG electrodes at 250 Hz. The data is a public 9-subject benchmark; each subject performed two sessions on different days, six runs per session, 48 trials per run, with the four classes evenly balanced.

Features. Deliberately not hand-engineered. The raw signal is bandpass filtered with a fifth-order Butterworth filter between 7 and 40 Hz — keeping the alpha and beta bands that carry motor-imagery information, and cutting the low frequencies where eye movement and blinks put most of their power. After that the network sees channels × time directly, and learns its own features. Doing feature extraction and classification in one model is the main design choice of the thesis.

Labels. One of four imagined movements: left hand, right hand, feet or tongue. That makes this a four-way classification problem. The model outputs a probability per class.

Hypothesis space. A convolutional neural network, using convolution and pooling to exploit the temporal and spatial structure across electrodes. A second variant adds transfer learning: weights from a pre-trained model are copied in and frozen, so fewer parameters need fitting and calibration time drops.

Loss function. Not confirmed. The full text is only available as a scanned PDF without a text layer, and neither the abstract nor the recorded defence states the training objective, so it is left unstated here rather than guessed.

What it found¶

Within-subject, with the model personalised to one person, mean accuracy was 58% ± 15%, which pre-training improved by a further 2.4% ± 1.3%. Between subjects, accuracy fell to 35% ± 11%.

Against the 25% expected from guessing on four balanced classes, 35% is better than guessing but far below usable, and the thesis says so. The gap between 58% and 35% is the main result, and it illustrates a general problem: the distribution shifts from person to person, so a model fitted to one brain transfers to the next at 35% accuracy. The weeks of calibration that motivated the work are reduced by pre-training, not eliminated.

Watch the defence¶

Full text¶

Read the thesis at Aaltodoc (open access)