[Paper Review] Applying Transfer Learning To Deep Learned Models For EEG Analysis

Axel Uran, Coert van Gemeren+3 more|arXiv (Cornell University)|Jul 2, 2019

EEG and Brain-Computer Interfaces24 references18 citations

TL;DR

This paper proposes a transfer learning approach for deep learning models in EEG signal classification, leveraging pre-trained features from large datasets to improve performance with limited labeled data. It achieves a 33% improvement over top traditional machine learning methods on the BCI Competition IV 2a dataset and demonstrates inter-experimental transfer learning that boosts performance by 18% on the 2b dataset, enabling robust classification with minimal data per subject or task.

ABSTRACT

The introduction of deep learning and transfer learning techniques in fields such as computer vision allowed a leap forward in the accuracy of image classification tasks. Currently there is only limited use of such techniques in neuroscience. The challenge of using deep learning methods to successfully train models in neuroscience, lies in the complexity of the information that is processed, the availability of data and the cost of producing sufficient high quality annotations. Inspired by its application in computer vision, we introduce transfer learning on electrophysiological data to enable training a model with limited amounts of data. Our method was tested on the dataset of the BCI competition IV 2a and compared to the top results that were obtained using traditional machine learning techniques. Using our DL model we outperform the top result of the competition by 33%. We also explore transferability of knowledge between trained models over different experiments, called inter-experimental transfer learning. This reduces the amount of required data even further and is especially useful when few subjects are available. This method is able to outperform the standard deep learning methods used in the BCI competition IV 2b approaches by 18%. In this project we propose a method that can produce reliable electroencephalography (EEG) signal classification, based on modest amounts of training data through the use of transfer learning.

Motivation & Objective

• To address the challenge of limited labeled EEG data in neuroscience by applying transfer learning to deep learning models.
• To improve EEG signal classification performance when only modest amounts of training data are available.
• To explore intra-experimental and inter-experimental transfer learning for knowledge transfer across subjects and experimental tasks.
• To outperform traditional machine learning methods in motor imagery EEG classification using deep learning with minimal data.
• To enable model reuse and generalization across different EEG experiments and paradigms through transferable feature representations.

Proposed method

• Pre-train a deep convolutional neural network (CNN) on a large EEG dataset to learn generic, low-level features.
• Fine-tune the pre-trained model on smaller target datasets from the BCI Competition IV 2a and 2b using transfer learning strategies.
• Implement two transfer learning variants: 'Frozen Learning' (freeze early layers) and 'Split Learning' (fine-tune only later layers) to optimize for limited data.
• Apply inter-experimental transfer by pre-training on one experiment and fine-tuning on another, even with different numbers of motor imagery classes.
• Use data augmentation and dropout to reduce overfitting during fine-tuning on small datasets.
• Evaluate performance using Cohen’s kappa (κ) and accuracy on held-out test sets from the BCI competition datasets.

Experimental results

Research questions

• RQ1 Can transfer learning improve deep learning performance on EEG classification with limited training data?
• RQ2 How does intra-experimental transfer learning (across subjects) enhance model generalization in motor imagery tasks?
• RQ3 Can inter-experimental transfer learning (across different experiments or tasks) further reduce data requirements?
• RQ4 How do different transfer learning strategies (e.g., frozen vs. split learning) compare to standard training and distributed learning?
• RQ5 To what extent can pre-trained models generalize across different EEG paradigms or numbers of motor imagery classes?

Key findings

• The proposed transfer learning method outperformed the top result from the BCI Competition IV 2a by 33% in terms of Cohen’s kappa score.
• The Frozen Learning method achieved the highest performance, closely followed by Split Learning, due to effective reuse of generic features from pre-trained layers.
• Inter-experimental transfer learning reduced data requirements and improved performance by 18% compared to standard deep learning approaches on the 2b dataset.
• Standard transfer learning improved performance by 5% on the four-movement task and 6% on the two-movement task compared to standard learning, though it underperformed Frozen and Split learning.
• The model showed strong transferability between motor imagery tasks, even when the number of classes differed between source and target tasks.
• The results suggest that sharing pre-trained EEG models across subjects and experiments can significantly enhance performance and reduce data dependency in BCI applications.