Work

A Bayesian machine learning framework for true zero-training brain-computer interfaces

Year: 2014
Type: dissertation
Abstract: Brain-Computer Interfaces (BCI) are developed to allow the user to take control of a computer (e.g. a spelling application) or a device (e.g. a robotic arm) by using just his brain signals. The concept of BCI was introduced in 1973 by Jacques Vidal. The early types of BCI relied on tedious user training to enable them to modulate their brain signals such that they can take control over the computer. Since then, training has shifted from the user to the computer. Hence, modern BCI systems rely on a calibration session, during which the user is instructed to perform specific tasks. The result of this calibration recording is a labelled data-set that can be used to train the (supervised) machine learning algorithm. Such a calibration recording is, however, of no direct use for the end user. Hence, it is especially important for patients to limit this tedious process. For this reason, the BCI community has invested a lot of effort in reducing the dependency on calibration data. Nevertheless, despite these efforts, true zero-training BCIs are rather rare. Event-Related Potential based spellers One of the most common types of BCI is the Event-Related Potentials (ERP) based BCI, which was invented by Farwell and Donchin in 1988. In the ERP-BCI, actions, such as spelling a letter, are coupled to specific stimuli. The computer continuously presents these stimuli to the user. By attending a specific stimulus, the user is able to select an action. More concretely, in the original ERP-BCI, these stimuli were the intensifications of rows and column in a matrix of symbols on a computer screen. By detecting which row and which column elicit an ERP response, the computer can infer which symbol the user wants to spell. Initially, the ERP-BCI was aimed at restoring communication, but novel applications have been proposed too. Examples are web browsing, gaming, navigation and painting. Additionally, current BCIs are not limited to using visual stimuli, but variations using auditory or tactile stimuli have been developed as well. In their quest to improve decoding performance in the ERP-BCI, the BCI community has developed increasingly more complex machine learning algorithms. However, nearly all of them rely on intensive subject-specific fine-tuning. The current generation of decoders has gone beyond a standard ERP classifier and they incorporate language models, which are similar to a spelling corrector on a computer, and extensions to speed up the communication, commonly referred to as dynamic stopping. Typically, all these different components are separate entities that have to be tied together by heuristics. This introduces an additional layer of complexity and the result is that these state of the art methods are difficult to optimise due to the large number of free parameters. We have proposed a single unified probabilistic model that integrates language models and a natural dynamic stopping strategy. This coherent model is able to achieve state of the art performance, while at the same time, minimising the complexity of subject-specific tuning on labelled data. A second and major contribution of this thesis is the development of the first unsupervised decoder for ERP spellers. Recall that typical decoders have to be tuned on labelled data for each user individually. Moreover, recording this labelled data is a tedious process, which has no direct use for the end user. The unsupervised approach, which is an extension of our unified probabilistic model, is able to learn how to decode a novel user’s brain signals without requiring such a labelled dataset. Instead, the user starts using the system and in the meantime the decoder is learning how to decode the brain signals. This method has been evaluated extensively, both in an online and offline setting. Our offline validation was executed on three different datasets of visual ERP data in the standard matrix speller. Combined, these datasets contain 25 different subjects. Additionally, we present the results of an offline evaluation on auditory ERP data from 21 subjects. Due to a less clear signal, this auditory ERP data present an even greater challenge than visual ERP data. On top of that we present the results from an online study on auditory ERP, which was conducted in cooperation with Michael Tangermann, Martijn Schreuder and Klaus-Robert Müller at the TU-Berlin. Our simulations indicate that when enough unlabelled data is available, the unsupervised method can compete with state of the art supervised approaches. Furthermore, when non-stationarity is present in the EEG recordings, e.g. due to fatigue during longer experiments, then the unsupervised approach can outperform supervised methods by adapting to these changes in the data. However, the limitation of the unsupervised method lies in the fact that while labelled data is not required, a substantial amount of unlabelled data must be processed before a reliable model can be found. Hence, during online experiments the model suffers from a warm-up period. During this warm-up period, the output is unreliable, but the mistakes made during this warm-up period can be corrected automatically when enough data is processed. To maximise the usability of ERP-BCI, the warm-up of the unsupervised method has to be minimised. For this reason, we propose one of the first transfer learning methods for ERP-BCI. The idea behind transfer learning is to share information on how to decode the brain signals between users. The concept of transfer learning stands in stark contrast with the strong tradition of subject-specific decoders commonly used by the BCI community. Nevertheless, by extending our unified model with inter-subject transfer learning, we are able to build a decoder that can decode the brain signals of novel users without any subject-specific training. Unfortunately, basic transfer learning models do perform as well as subject-specific (supervised models). For this reason, we have combined our transfer learning approach with our unsupervised learning approach to adapt it during usage to a highly accurate subject-specific model. Analogous to our unsupervised model, we have performed an extensive evaluation of transfer learning with unsupervised adaptation. We tested the model offline on visual ERP data from 22 subjects and on auditory ERP data from 21 subjects. Additionally, we present the results from an online study, which was also performed at the TUBerlin, where we evaluate transfer learning online on the auditory AMUSE paradigm. From these experiments, we can conclude that transfer learning in combination with unsupervised adaptation results in a true zero training BCI, that can compete with state of the art supervised models, without needing a single data point from a calibration recording. This method allows us to build a BCI that works out of the box.
Language: English
Open Access status: green
OpenAlex ID: W2161256981
Cites: 124
Cited by: 4
Related to: 10