Calibration-free online test-time adaptation for electroencephalography motor imagery decoding (2024)
Open in webOpen in zoteroOpen pdf
1 Abstract
Providing a promising pathway to link the human brain with external devices, Brain-Computer Interfaces (BCIs) have seen notable advancements in decoding capabilities, primarily driven by increasingly sophisticated techniques, especially deep learning. However, achieving high accuracy in real-world scenarios remains a challenge due to the distribution shift between sessions and subjects. In this paper we will explore the concept of online test-time adaptation (OTTA) to continuously adapt the model in an unsupervised fashion during inference time. Our approach guarantees the preservation of privacy by eliminating the requirement to access the source data during the adaptation process. Additionally, OTTA achieves calibration-free operation by not requiring any session- or subject-specific data. We will investigate the task of electroencephalography (EEG) motor imagery decoding using a lightweight architecture together with different OTTA techniques like alignment, adaptive batch normalization, and entropy minimization. We examine two datasets and three distinct data settings for a comprehensive analysis. Our adaptation methods produce state-of-the-art results, potentially instigating a shift in transfer learning for BCI decoding towards online adaptation.
2 NOTES
In this paper the authors propose an online test-time adaptation (OTTA) strategy, so that a model trained on one (or multiple) subjects can be fine-tuned for another in a situation similar to what is expected to happen in a real scenario. The main idea is that every sample from the test set should could in sequence and one at a time. Essentially, this means a batch size of 1. They used Batch Normalization, Riemannian Alignment, Euclidean Alignment, all of which are techniques that require a batch size bigger then 1. So, to solve it they create a buffer. As each sample is inputted into to the model it goes to a buffer following a first in first out order, up to a maximum limit (set in here to 32 and in @junqueiraSystematicEvaluationEuclidean2024 to 24). This allows their alignment to be adapted over time, and also use Adaptive Batch Normalization. It is interesting to notice that instead of simply using something like Incremental Euclidean/Riemannian Alignment they opted instead to use exponential weighting for the mean, which makes total sense, as the reference matrix should probably focus on more recent trials and is also the smartest choice to keep up with the idea of buffering trials. Their best results used Riemannian Alignment (with Exponential Moving Average), Batch Normalization (with only the mean and std from the target samples) and also label smoothing (with for cross-session and for cross-subject). Surprisingly the label smoothing had a substantial impact on accuracy, improving by on the cross-subject settings over the authors previous best result. Overall a great paper, where the best idea is certainly the use of a buffer to obtain statistics from the test data.