Towards Discriminant Analysis Classifiers Using Online Active Learning via Myoelectric Interfaces

Andres G. Jaramillo-Yanez, Marco E. Benalcázar, Sebastian Sardiña, Fabio ZambettaAAAI 2022 · DOI

Introduction

Myoelectric interfaces, which utilize surface electromyography (sEMG) signals to control prosthetic limbs and wearable devices, face a significant challenge: signal non-stationarity. Factors such as muscle fatigue, electrode displacement, and skin impedance changes cause "covariate shift," where the statistical properties of the input signals change over time. Traditionally, this requires users to undergo frequent, tedious re-training sessions to maintain accuracy. This paper presents an online Discriminant Analysis (DA) framework that adapts to these shifts in real-time, reducing training fatigue and improving long-term usability.

Methodology

The proposed system transitions from traditional batch learning to an Online Active Learning paradigm, characterized by three core components:

  1. Incremental DA Updates: The authors derived a method to update the mean vectors and covariance matrices of Linear and Quadratic Discriminant Analysis (LDA/QDA) classifiers sample-by-sample. They mathematically proved that this incremental update yields the exact same model parameters ($\Theta_t = \Theta^\star_t$) as a batch model trained on the entire accumulated dataset.
  2. Soft-Labeling with Contribution Weighting: To enable learning without constant user feedback, the system employs pseudo-labels. A novel soft-labeling technique calculates a "contribution weight" for each new data point based on its distance to class centroids.
  3. Entropy-Based Robustness: To prevent the model from learning from its own errors, the system uses Shannon entropy and the Matthews Correlation Coefficient (MCC). Samples with high uncertainty (high entropy) are assigned lower weights, minimizing their influence on the model's parameters.

Robustness via Weighted Covariance

A key technical insight is the use of a "Weighted" covariance update. In traditional DA, mislabeled data can cause class distributions to overlap significantly. By calculating covariance relative to individual gesture means rather than a global mean, the proposed approach maintains better class separation even in the presence of noise.

Comparison of traditional (a) and weighted (b) class distribution updates. The weighted approach (b) maintains better class separation during online updates.

Experimental Results

The framework was validated across five diverse sEMG datasets: NinaPro5, Capgmyo_dbb, MyoArmband, Long-Term 3DC, and EMG-EPN-120. These datasets represent a wide range of hardware, user populations (up to 612 subjects), and temporal signal variations.

Performance and Efficiency

The online classifier consistently outperformed the initial baseline (trained on minimal data) across all datasets. Statistical analysis using the Friedman test ranked the proposed soft-labeling technique as the top-performing method for QDA and tied for first for LDA, surpassing established methods like Nigam-based soft-labeling.

Classification accuracy across five benchmark datasets. The proposed online classifier using pseudo-labels (red) significantly improves upon the initial baseline (grey) and competes with the fully labeled upper bound (blue).

Crucially, the system operates with $O(1)$ time complexity relative to the data stream. This ensures that the computational lag remains constant regardless of how long the device has been in use, making it ideal for real-time applications.

Conclusion

By combining incremental mathematical equivalence with robust soft-labeling, this work provides a path toward myoelectric interfaces that "learn on the job." The system effectively mitigates covariate shift while maintaining the computational efficiency required for wearable hardware. Future work will address class imbalance issues and the sensitivity of the model to the quality of the initial "seed" training data.

Summary updated

BibTeX
@inproceedings{jaramilloyanez2022towards,
  author = {Andres G. Jaramillo-Yanez and Marco E. Benalcázar and Sebastian Sardiña and Fabio Zambetta},
  title = {{Towards Discriminant Analysis Classifiers Using Online Active Learning via Myoelectric Interfaces}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i6.20658},
}
Figure