PaperlayerView exact revision on arXiv

Source-linked AI summary

Transfer Learning in Brain-Computer Interfaces

Vinay Jayaram, Morteza Alamgir, Yasemin Altun, Bernhard Schölkopf, Moritz Grosse-Wentrup

arXiv:1512.00296v1cs.HCq-bio.NC

TL;DR

BCI performance benefits from more training data, but variation across subjects and sessions limits transfer. The paper presents a general transfer-learning framework and an EEG-specific regression method, achieving comparable or better accuracy with fewer trials.

  • Problem

    Variation in training-data distributions across subjects and sessions, together with few trials per patient session, limits transferability in BCIs.

  • Method

    The paper reviews BCI transfer-learning techniques and introduces a feature-space-independent framework plus a regression method designed for EEG structure.

  • Results

    The proposed methods better handle session-to-session and subject-to-subject variability than simple pooling, achieving accuracies comparable to or better than single-session training with far fewer trials.

  • Takeaways & Limitations

    Transfer learning can support BCI training across subjects and sessions while reducing the number of trials needed relative to single-session training.

  • Takeaways & Limitations

    The proposed regression method has undesirable features, including sensitivity to smart initialization of spatial weights.

Abstract

from arXiv · show

The performance of brain-computer interfaces (BCIs) improves with the amount of available training data, the statistical distribution of this data, however, varies across subjects as well as across sessions within individual subjects, limiting the transferability of training data or trained models between them. In this article, we review current transfer learning techniques in BCIs that exploit shared structure between training data of multiple subjects and/or sessions to increase performance. We then present a framework for transfer learning in the context of BCIs that can be applied to any arbitrary feature space, as well as a novel regression estimation method that is specifically designed for the structure of a system based on the electroencephalogram (EEG). We demonstrate the utility of our framework and method on subject-to-subject transfer in a motor-imagery paradigm as well as on session-to-session transfer in one patient diagnosed with amyotrophic lateral sclerosis (ALS), showing that it is able to outperform other comparable methods on an identical dataset.

1 INTRODUCTION

BCI training data and decoding rules vary across subjects and sessions, limiting direct reuse. The paper reviews transfer-learning approaches and distinguishes domain adaptation from rule adaptation.

  • Motivation: BCI performance can change substantially across sessions and subjects, even when the same classifier is used.The EEG signal is nonstationary, so each trial can represent a slightly new task.
  • Motivation: Pooling recordings can help, but distribution differences across subjects and sessions make its effectiveness variable.These differences motivate explicitly modeling shared structure across datasets.
  • Contribution: The paper presents a unified BCI transfer-learning approach and evaluates it for motor-imagery subject transfer and ALS session transfer.The proposed method is designed to handle poorly performing subjects and extremely high-dimensional feature spaces.
  • Transfer-learning approaches: Domain adaptation seeks an invariant data representation where one decision rule classifies samples across datasets.Many BCI methods use invariant projections or feature spaces, including CSP-based approaches.
  • Transfer-learning approaches: Rule adaptation models how classification rules differ, reducing the search space for learning decision boundaries on new datasets.This differs from transforming all data into one shared invariant space.
  • Previous work: Prior transfer methods reduced calibration time or preserved performance by learning shared spatial filters, classifiers, or covariance-based representations.Reported approaches include clustering prototypical filters, sparse cross-subject filter subsets, and invariant covariance projections.

2 A GENERAL FRAMEWORK FOR TRANSFER LEARNING IN BCIS

The framework treats each subject or session as a related prediction task and learns shared structure alongside task-specific models. It applies across spatiotemporal feature spaces and introduces a BCI-specific feature-decomposition method.

  • Framework overview: The framework is applicable to any spatiotemporal feature space and supports multi-session and multi-subject transfer equally.It also introduces a BCI-specific method for reducing feature-space dimensionality.
  • Preliminaries: Each subject or recording session is modeled as one task with task-specific weights that predict its stimulus labels.The framework can also be applied to regression problems.
  • Preliminaries: The model uses EEG log-bandpower features across scalp locations, while the framework can also use event-related-potential timepoints.With E electrodes and F frequency estimates, the feature count is d = E × F.
  • Independent training: Independent training estimates each subject/session model by minimizing squared prediction error, with regularization controlling model complexity.The loss is represented using the Euclidean norm ∥Xsws − ys∥2.
  • Joint training: Joint training learns task models together with a shared mean and covariance that encode subject/session-invariant characteristics.The shared structure acts as prior information, while deviations from the shared mean represent task-specific characteristics.
  • Feature decomposition: Feature decomposition reduces the decoding-model parameter count from EF to E + F.The decomposition operates on a feature matrix X ∈ R^E×F for each trial.

3 EXPERIMENTS

The experiments evaluate transfer learning in two real-world settings: subject-to-subject motor imagery and session-to-session neurofeedback in a single patient with ALS.

  • Subject-to-subject transfer: The first experiment evaluates subject-to-subject transfer using motor-imagery data from ten healthy subjects.It compares the initial multitask-learning algorithm with a version that decomposes spectral and spatial features.
  • Session-to-session transfer: The second experiment evaluates session-to-session transfer in a neurofeedback paradigm recorded from one patient with ALS.This experiment uses the feature-decomposition version, referred to as FD.

Paradigm

The motor-imagery paradigm instructed participants to imagine left- or right-hand movement during timed trials cued by an on-screen arrow.

  • Trial structure: Each trial began with a three-second pause before an arrow instructed left- or right-hand haptic motor imagery.The arrow direction indicated which hand participants should imagine moving.
  • Trial structure: After seven additional seconds, the arrow disappeared, marking the end of the trial and signaling participants to stop imagining movement.

Dataset

The study included ten healthy subjects who performed left- and right-hand motor-imagery trials while EEG was recorded with 128 channels. Each subject contributed 150 trials per class in pseudorandomized order without feedback.

  • Ten healthy subjects participated, including two females with a mean age of 25.6 ± 2.5 years.
  • One subject had prior motor-imagery experience, whereas the others were naïve to motor imagery and BCIs.
  • EEG was recorded from 128 channels using the extended 10-20 system, with Cz as reference, at 500 Hz.
  • A temporal analog high-pass filter with a 10-second time constant was used with BrainAmp amplifiers.
  • Each subject completed 150 trials per class for left- and right-hand motor imagery.
  • Trials were presented in pseudorandomized order, and subjects received no feedback during the experiment.

Feature Extraction

EEG features were constructed using data-independent spatial filtering, frequency-band decomposition, and log-bandpower measurements. The resulting representation had 128 × 12 dimensions per trial.

  • EEG recordings were first spatially filtered with a surface Laplacian setup.
  • More sophisticated spatial filters such as CSP and beamforming were excluded to keep the spatial-filtering setup data-independent.
  • For each subject, trial, and electrode, 2-Hz frequency bands spanning 7–29 Hz were extracted.
  • A discrete Fourier transform with a Hanning window was computed over each trial.
  • Log-bandpower over the final seven seconds of each trial formed the 128 × 12-dimensional feature vector.

Classification Performance

Classification was evaluated for subject-to-subject transfer using pooled and subject-specific controls against multitask methods. MT FD was generally strongest, while multitask learning retained high-data accuracy comparable to subject-specific training and produced weights consistent with motor-related EEG signals.

  • Controls included standard and FD ridge regression and a linear-kernel SVM, evaluated with pooled or test-subject-only training data.
  • MT FD consistently outperformed the other algorithms at nearly all levels of test-subject data.
  • With 200 training trials, MT FD achieved a higher mean accuracy than every other algorithm by using subject-specific data more effectively.
  • In pooled training, MT FD matched pooled-data accuracy with few trials and both multitask methods improved beyond pooled controls with more trials.
  • MT without FD initially underperformed pooled ridge regression because each subject had few trials relative to the feature space, with 300 trials versus 1400 features.
  • MT without FD used subject-specific data more effectively than pooling algorithms, as indicated by its higher classification-curve slope.
  • At high trial counts, MT accuracies were nearly identical to subject-specific training, indicating no reduction in maximum achievable accuracy for MT.
  • MT FD weights emphasized motor-cortex electrodes, opposite cortical signs, and the μ band, consistent with expected brain-derived motor-imagery signals.

3.2 Session-to-Session Transfer

The session-to-session transfer experiment addressed the low number of trials common in BCI paradigms, especially for patient populations. The authors applied the FD approach to a 30-session dataset with very limited training and testing data per condition.

  • Low numbers of trials per session are a common issue in BCI paradigms, particularly those involving patient populations.
  • The FD approach was evaluated on a 30-session dataset.
  • Each session contained only ten training trials and 10–20 test trials for each condition.

Data Collection

The ALS participant trained over thirty sessions across fifteen months to modulate 1–5 Hz power in the precuneus. EEG was recorded with 121 channels and processed using a standardized montage and reference.

  • The ALS participant trained to modulate 1–5 Hz δ-bandpower in the precuneus over thirty sessions spanning fifteen months.
  • The participant’s ALS-FRS-R score decreased from 33 to 9 during the fifteen-month study period.
  • Each sixty-second trial required the participant to deviate up or down from a session-specific mean, with twenty trials per run.
  • The protocol used two to three runs per session, with the first session dedicated entirely to training.
  • Recordings used a 121-channel EEG system sampled at 500 Hz, with electrodes placed under the extended 10–20 system and converted to common-average reference.
  • Figure 4 summarizes learned multitask priors across frequency windows and spatial weights over training-subject groups, including motor-cortex concentration.

Feature extraction and training

The study extracted frequency-domain EEG features, updated classifiers in a pseudo-online manner, and compared multitask, pooled ridge-regression, and SVM approaches across sessions. Multitask learning and pooled ridge regression achieved the strongest overall session-level performance, with multitask learning offering a higher minimum accuracy.

  • Feature extraction: Independent component analysis with SOBI removed artifacts, and selected cortical components were re-projected into electrode space.
  • Feature extraction: Log-powers across seven frequency bands produced a 121 × 7 feature space for each trial.
  • Training procedure: The first session trained the classifier; subsequent sessions used their first run for updates before classifying the remaining runs pseudo-online.
  • Training procedure: Between sessions, Algorithm 2 was rerun using all data from the most recent session because the non-initialized case performed better on these data.
  • Results: Multitask learning and pooled ridge regression had the highest median classification accuracy, at 85%, across sessions.
  • Results: Both methods significantly outperformed single-session ridge regression (p < 0.0001, Wilcox signed-rank test).
  • Results: Multitask learning achieved a higher minimum classification accuracy than pooled regression, 65% versus 60%.
  • Results: Learned weights concentrated above the precuneus, with selected strong channels rather than a smooth topography, consistent with feature-selection regularization.

4 DISCUSSION

The framework supports transfer learning across subjects and sessions, works across feature spaces, and can reduce feature dimensionality and training-trial requirements. Its regression method performs well but depends on initialization, adds computational considerations, and has scope and scalability limitations.

  • Contributions: The framework provides a unified approach to transfer learning across subjects and sessions and applies across feature spaces.It is presented as complementary to domain adaptation methods.
  • Results: Feature decomposition reduces the feature-space dimension and can require fewer training trials than other classification methods.The method was reported to outperform an SVM with an equal feature space in zero-training and within-session transfer settings.
  • Limitations: Initialization strongly affects the proposed regression method, with poor initialization producing very poor motor-imagery results and potentially failing to find predictive features.The authors report no clear rule for when non-informative or ridge-regression initialization is preferable and recommend testing both empirically.
  • Computational considerations: Feature decomposition adds an iterative loop, although the FD case converged orders of magnitude faster and was faster overall than the non-FD case in practice.The alternative regular regression requires inversion of a much larger matrix for every session.
  • Limitations: The approach is limited to one spatial and frequency weight vector, preventing simultaneous use of multiple signals, and must be rerun after adding subjects or sessions.The full iterative rerun becomes inefficient as the number of subjects or sessions increases.
  • Method interpretation: The multitask formulation models shared structure through a distribution over task-specific weight vectors and can be added as a convex loss term.The inferred prior parameters also provide spatial and spectral topographies associated with the task.

5 CONCLUSION

The paper presents transfer-learning methods for BCI data across subjects and sessions, extending knowledge transfer into arbitrary spatiotemporal feature spaces. Experiments indicate improved handling of variability compared with simple pooling, while retaining performance with fewer training trials.

  • Previous BCI transfer-learning approaches focused mainly on spatial filtering and domain adaptation, leaving knowledge transfer within feature spaces underexplored.
  • The proposed method transfers knowledge from previous subjects to new ones in any desired spatiotemporal feature space and can complement other paradigms.
  • Testing covered motor imagery and a novel cognitive paradigm, including both session-to-session and subject-to-subject variability.
  • The proposed methods handled session-to-session and subject-to-subject variability better than simple pooling, achieving accuracies comparable to or better than single-session training with far fewer training trials.
  • The framework supports alternative objective functions for determining priors over decision boundaries that minimize other types of error.
Loading 1512.00296v1…