(paper 65) SeqCLR

Contrastive Representation Learning for EEG Classification

Contents

1. Abstract
2. Introduction
3. Method
   1. Channel Recombination & Preprocessing
   2. Channel Augmentations
   3. Learning Algorithm

0. Abstract

Interpreting and labeling EEG : challenging

SeqCLR ( Sequential Contrastive Learning of Representations )

Present a framework for learning representations from EEG signals via contrastive learning

• recombine channels from multi-channel recordings

  \(\rightarrow\) increase the number of samples quadratically per recording.

• train a channel-wise feature extractor

  • by extending the SimCLR to TS

• introduce a set of augmentations for EEG

1. Introduction

Presenting a new framework that allows us to ..

• (1) combine multiple EEG datasets
• (2) use the underlying physics of EEG signals to multiply the number of samples (quadratic increase)
• (3) learn representations in a self-supervised manner via CL

Details

• Modify the SimCLR framework for TS
• In contrast to images, not clear what augmentations could be beneficial for TS
  • consulted EEG researchers to select a set of transformations

2. Method

(1) Channel Recombination & Preprocessing

we obtain \(n \times (n-1) + n = n^2\) new channels for \(n\)-channel recording

(2) Channel Augmentations

A key ingredient of CL = augmentations

we chose the transformations as…

Strength of each transformation :

(3) Learning Algorithm

SeqCLR ( Sequential Contrastive Learning of Representations )

• like SimCLR, contains 4 modules

1. Channel Augmenter

   • for each channel, the module randomly applies 2 augmentations
   • \(N \rightarrow 2N\) augmented channels

2. Channel Encoder

   • transforms an input channel into 4 feature channels of same length

   • enables us to encode sequences of different lengths for different downstream tasks

   • designed 2 encoder :

     • (1) A recurrent encoder

       • with a multi-scale input ( using down & up sampling of the channel )
       • uses 2 recurrent residual units

     • (2) A convolutional encoder

       • utilizes reflection paddings

         ( to ensure the output signal is of the same length as the input signal )

       • uses 4 convolutional residual units

3. Projector

   • recurrent projection head

   • collapses the output of the encoder into 32-dim
   • uses downsampling & bidirectional LSTM units

4. Contrastive Loss

   • identical to NT-Xent ( in SimCLR )
   • \(\ell_{i, j}=-\log \frac{\exp \left(\operatorname{sim}\left(\boldsymbol{z}_i, \boldsymbol{z}_j\right) / \tau\right)}{\sum_{k \neq i}^{2 N} \exp \left(\operatorname{sim}\left(\boldsymbol{z}_i, \boldsymbol{z}_k\right) / \tau\right)}\).

5. Classifier

   • ( for downstream cls task ) discard the projector & use classifier
   • details
     • output dim = # of classes
     • log softmax