PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Transfer Learning for EEG Signal Classification

Ce Ju, Dashan Gao, Ravikiran Mane, Ben Tan, Yang Liu, Cuntai Guan

arXiv:2004.12321v5cs.LGeess.SP

TL;DR

Small EEG datasets and privacy restrictions make joint model training across sources difficult, while device differences also create heterogeneous feature spaces. The paper proposes federated transfer learning using covariance-based neural networks for EEG classification. In subject-specific analysis, it reports 6% better classification accuracy than two state-of-the-art DL architectures.

  • Problem

    EEG-BCI datasets are small and scattered, and privacy concerns restrict combining them for joint model training; device differences also create heterogeneous feature spaces with limited labels.

    Source: §I · p. 1

  • Method

    FTL uses spatial covariance matrices as inputs to a neural BCI model adapted for cross-subject learning with domain adaptation and a federated framework.

    Source: §I · p. 1, §III · p. 2

  • Results

    6% better classification accuracy than the other two state-of-the-art DL architectures in subject-specific analysis.

    Source: C. Discussion · p. 5

  • Takeaways & Limitations

    The authors conclude that federated learning can train BCI models across heterogeneous datasets and extract knowledge from heterogeneous BCI data.

    Source: §V · p. 5

    Evaluation is reported on the PhysioNet EEG Motor Movement/Imagery Dataset, for subject-specific and subject-adaptive analysis.

    Source: §I · p. 1

Abstract

from arXiv · show

The success of deep learning (DL) methods in the Brain-Computer Interfaces (BCI) field for classification of electroencephalographic (EEG) recordings has been restricted by the lack of large datasets. Privacy concerns associated with EEG signals limit the possibility of constructing a large EEG-BCI dataset by the conglomeration of multiple small ones for jointly training machine learning models. Hence, in this paper, we propose a novel privacy-preserving DL architecture named federated transfer learning (FTL) for EEG classification that is based on the federated learning framework. Working with the single-trial covariance matrix, the proposed architecture extracts common discriminative information from multi-subject EEG data with the help of domain adaptation techniques. We evaluate the performance of the proposed architecture on the PhysioNet dataset for 2-class motor imagery classification. While avoiding the actual data sharing, our FTL approach achieves 2% higher classification accuracy in a subject-adaptive analysis. Also, in the absence of multi-subject data, our architecture provides 6% better accuracy compared to other state-of-the-art DL architectures.

I. INTRODUCTION

EEG-BCI research faces small, scattered datasets and privacy barriers to combining them; device heterogeneity also leaves many clients with limited labeled data. The paper proposes FTL, combining covariance-based neural methods with federated transfer learning.

I. INTRODUCTION

  • Privacy concerns and the cost of collecting EEG data leave small datasets scattered across sources, motivating joint analysis without sharing the data.The introduction notes that EEG data may expose brain activity and cites regulations that restrict exchange without user approval.Source: §I · p. 1
  • Heterogeneous devices produce different feature spaces, while many clients have limited labels, leaving no suitable source domain for standard source-target adaptation.The authors say existing heterogeneous domain-adaptation solutions assume multiple source domains with abundant labels and one target domain with limited labels.Source: §I · p. 1
  • The proposed Federated Transfer Learning (FTL) adapts a covariance-based neural BCI method to cross-subject learning using a federated framework.The model takes spatial covariance matrices as input and uses domain adaptation for subject-adaptive analysis.Source: §I · p. 1
  • The paper evaluates subject-specific and subject-adaptive analysis on the PhysioNet EEG Motor Movement/Imagery Dataset and reports matching state-of-the-art accuracy without sharing EEG data.The introduction lists these analyses and privacy-preserving comparison among the paper’s contributions.Source: §I · p. 1

1) Covariance Methods in BCI:

III. METHODOLOGY

FTL uses spatial covariance matrices as inputs and processes them through manifold reduction, a common embedded space, tangent projection, and a federated classification layer. The paper distinguishes this transfer architecture from a non-transfer version and assumes filtered EEG signals with covariance matrices on embedded SPD submanifolds.

III. METHODOLOGY

  • A short-time EEG trial X is represented by its E × E spatial covariance matrix S, an SPD matrix, before entering the four-layer architecture.The layers are manifold reduction, common embedded space, tangent projection, and federated classification.Source: §III · p. 2
  • Manifold reduction maps high-dimensional SPD manifolds to lower-dimensional ones, and the common embedded space provides a shared low-dimensional SPD manifold for transfer learning.Tangent projection maps matrices on SPD manifolds to a tangent space, a local linear approximation of the curved space.Source: §III · p. 2
  • The federated layer classifies the projected inputs, with its neural-network parameters updated by federated aggregation in the transfer setting.The transfer architecture uses M, C, T, and F; the non-transfer architecture uses only M and T.Source: §III · p. 2

A. Notation

  • The setup assumes band-pass-filtered EEG and models each subject’s spatial covariance matrices as distributed on an embedded SPD submanifold.The number of trials is denoted Ni for subject i.Source: A. Notation · p. 2

B. Manifold Reduction Layer

The manifold reduction layer learns subject-specific mappings from high-dimensional SPD manifolds to a shared lower-dimensional SPD manifold. The common embedded space is encouraged to bring different subjects’ reduced covariance distributions close together using MMD.

B. Manifold Reduction Layer

  • For each subject, a neural-network reduction maps spatial covariance matrices from that subject’s SPD manifold to the common SPD manifold N.The reduction is approximated by a neural network and constructed as a bilinear mapping.Source: B. Manifold Reduction Layer · p. 2

C. Common Embedded Space

  • Maximum mean discrepancy measures distances between subjects’ reduced covariance distributions on N, with smaller MMD corresponding to closer reduced matrices.The method expresses the distributions through a feature map into a reproducing-kernel Hilbert space.Source: C. Common Embedded Space · p. 2

D. Tangent Projection Layer

The tangent projection layer maps reduced covariance matrices from the curved SPD manifold to a tangent space at a reference matrix. With the identity reference, the paper gives a compact matrix-logarithm expression for the tangent vector.

D. Tangent Projection Layer

  • The logarithmic map projects a spatial covariance matrix onto the tangent space of a reference matrix, providing a local linear representation of the curved SPD space.The tangent space is associated with a reference matrix P on the common embedded SPD manifold.Source: D. Tangent Projection Layer · p. 2
  • Using the identity reference matrix, the tangent projection layer maps each reduced covariance matrix X to log(X).The paper identifies this as a compact expression for the tangent vector.Source: D. Tangent Projection Layer · p. 3

E. Federated layer

The federated layer uses neural-network classifiers on a common tangent-space representation and federated averaging to aggregate model updates across subjects.

E. Federated layer

  • Each subject’s classifier uses the same architecture, including fully connected and flatten layers and an activation function, on the Euclidean tangent space of the common manifold.The common manifold’s tangent space is locally homeomorphic to a Euclidean space of dimension d.Source: E. Federated layer · p. 3
  • Federated averaging aggregates feature maps over each subject’s clients and classifier updates over all clients, then sends the updated global model back to clients.Clients continue training with their local data on the reduced common manifold until convergence.Source: E. Federated layer · p. 3

F. Architecture and Loss

The architecture distinguishes non-transfer and transfer versions through their loss objectives: the former uses classification loss alone, while the latter also includes domain loss.

F. Architecture and Loss

  • The non-transfer version combines repeated manifold reduction and tangent projection layers with neural networks, using classification loss alone.The subscript i represents each subject.Source: F. Architecture and Loss · p. 3
  • The transfer version adds domain loss to classification loss; classification uses cross-entropy, while domain terms use MMD with preset weights.The cross-entropy compares predicted labels with ground truth; MMD measures distances between distributions using a preset feature map.Source: F. Architecture and Loss · p. 3

IV. EXPERIMENTS

The experiments use PhysioNet EEG motor-imagery data, focusing on imagined left- and right-hand movements across 109 subjects.

A. Experimental Setup

  • The evaluation uses the PhysioNet EEG Motor Imagery dataset.The experiments assess subject-specific and subject-adaptive analyses.Source: A. Experimental Setup · p. 3

1) MI Dataset Description:

  • The dataset contains 2-class imagined left- and right-hand movement trials from 109 subjects, recorded with 64 electrodes at 160 Hz.Figure 2 plots average classifier accuracy against subject number from 1 to 109.Source: 1) MI Dataset Description · p. 3, Figure 2 · p. 4

2) Evaluated Algorithms:

The experiments compare FTL with subspace, covariance, and deep-learning classifiers and report cross-validation accuracy across subject-specific and subject-adaptive settings.

2) Evaluated Algorithms:

  • The comparison includes CSP with LDA or SVM and FBCSP as subspace-method baselines.CSP uses an LDA or SVM classifier; FBCSP is described as the best-performing CSP method for motor-imagery classification.Source: 2) Evaluated Algorithms · p. 4
  • Covariance and deep-learning comparators include MDM, TSM, R-Kernel, EEGNet, and DeepConvNet.These include SPD-manifold distance and tangent-space classifiers, a Riemannian-kernel method, and two neural architectures.Source: 2) Evaluated Algorithms · p. 4

3) Architecture of FTL in Experiments:

  • The subject-specific FTL architecture uses three manifold-reduction layers, a tangent projection layer, and a federated layer without federated aggregation.Its learning rate starts at 0.1, decays after 50 epochs, and training stops when loss falls below 0.1.Source: 3) Architecture of FTL in Experiments · p. 4
  • Subject-adaptive FTL uses 27 good subjects as the source and one bad subject as the target, with a shared 4 × 4 SPD common space and federated aggregation.The source uses three manifold-reduction layers and the target uses two; the federated layer is fully connected with dimensions 16 × 2.Source: 3) Architecture of FTL in Experiments · p. 4
  • Table I reports cross-validation average accuracy for nine classifiers over 109 subjects; Table II covers three settings, including subject-specific averages for 27 good and 28 bad subjects.The supplied table descriptions do not state individual accuracy values.Source: Table II · p. 4, Table I · p. 4

B. Experimental Results and Analysis

In subject-specific 2-class motor imagery analysis, FTL outperforms two state-of-the-art deep-learning methods by combining covariance methods with deep learning. The section also defines the good- and bad-subject groups used in the subsequent adaptive analysis.

B. Experimental Results and Analysis

  • Experiments used 5-fold cross-validation, with four folds for training and one for testing, and Figure 2 showed no competing classifier outperforming the others.The same random fold assignment was maintained across classification methods.Source: B. Experimental Results and Analysis · p. 4

1) Subject-specific Analysis for 2-class MI Classification:

2) Subject-adaptive Analysis for 2-class MI Classification:

FTL achieved 6% better subject-specific classification accuracy than two state-of-the-art DL architectures and the best accuracy in subject-adaptive analysis. The paper reports that federated learning can support distributed training across heterogeneous BCI datasets.

2) Subject-adaptive Analysis for 2-class MI Classification:

  • In the transfer setting, each bad subject's training data combined data from good subjects with four folds of that bad subject's data, while the remaining fold was used for testing.Modeling with good-subject signals yielded better accuracy than modeling with bad-subject data alone, and FTL was effective in the transfer setting.Source: 2) Subject-adaptive Analysis for… · p. 5

C. Discussion

  • 6% better classification accuracy than the other two state-of-the-art DL architectures was achieved in subject-specific analysis; FTL also achieved the best accuracy in subject-adaptive analysis.The discussion attributes the subject-specific result to covariance-based EEG representation and identifies domain adaptation as useful in the adaptive analysis.Source: C. Discussion · p. 5

V. CONCLUSIONS

  • The experiments support the feasibility of federated learning for distributed BCI model training across datasets with heterogeneous configurations.The conclusion also reports the potential of extracting knowledge from heterogeneous BCI data.Source: §V · p. 5
Loading 2004.12321v5…