Targeted Adversarial Denoising Autoencoders (TADA) for Neural Time Series Filtration

TL;DR

The paper introduces Targeted Adversarial Denoising Autoencoders (TADA) for neural time-series filtration, addressing limitations of large, slow DL denoisers by combining a meta-targeting LSTM-CNN ensemble, a lightweight convolutional denoising autoencoder, adversarial training, and a covariance-driven logistic scale targeting layer. The approach achieves competitive reconstruction metrics with a compact parameter budget (<400,000) and fast training times, enabling potential real-time deployment. Key contributions include a novel meta-targeting strategy for SNR adaptation, a CC-based loss with spectral preservation, and a robust scale-targeting module that preserves correlation while reducing error. This framework offers a practical path toward real-time EMG artifact removal in neural interfaces and can be extended to multi-channel settings, pending broader validation.

Abstract

Current machine learning (ML)-based algorithms for filtering electroencephalography (EEG) time series data face challenges related to cumbersome training times, regularization, and accurate reconstruction. To address these shortcomings, we present an ML filtration algorithm driven by a logistic covariance-targeted adversarial denoising autoencoder (TADA). We hypothesize that the expressivity of a targeted, correlation-driven convolutional autoencoder will enable effective time series filtration while minimizing compute requirements (e.g., runtime, model size). Furthermore, we expect that adversarial training with covariance rescaling will minimize signal degradation. To test this hypothesis, a TADA system prototype was trained and evaluated on the task of removing electromyographic (EMG) noise from EEG data in the EEGdenoiseNet dataset, which includes EMG and EEG data from 67 subjects. The TADA filter surpasses conventional signal filtration algorithms across quantitative metrics (Correlation Coefficient, Temporal RRMSE, Spectral RRMSE), and performs competitively against other deep learning architectures at a reduced model size of less than 400,000 trainable parameters. Further experimentation will be necessary to assess the viability of TADA on a wider range of deployment cases.

Figures (9)

• Figure 1: The architecture of the TADA system, with the targeting system depicted on the left and the adversarial autoencoder process on the right. The incoming signal is routed through the meta-targeting ensemble (upper left) before being passed through the autoencoder (right) and then through the covariance-driven rescale targeting function (lower left). The loss function (upper right) and the adversarial discriminator (lower right) assist in training. More extensive details are included in the Appendix.
• Figure 2: Two examples (one left, one right) of the logistic target rescaling workflow leveraging low noise segment identification and weighting. Contaminated signals (A) are routed through the autoencoder, and the subsequent scale targeting process (B) maps the AE output to the appropriate final range (C). Ground truth is depicted in the bottom panel (D) for reference. More precise formulations are included in the Appendix.
• Figure 3: Two examples of TADA filtration demonstrating system noise reduction capabilities (spike artifact on the left; continuous artifact on the right). Panels (A), (B), and (C) depict contaminated, ground truth, and the TADA-reconstructed signal, respectively. TADA time series filtration yields improved correlation with ground truth (D) and reduced mean squared error (E).
• Figure 4: The correlation-driven autoencoder loss function.
• Figure 5: CC, TRRMSE, and SRRMSE test performance distribution on the -7 dB, -2.5 dB, and 2 dB test cases. The upper panels depict performance scatterplots of the individual test samples, with CC, TRRMSE, and SRRMSE denoted on the three axes. The lower panels depict histograms (from top to bottom) of TRRMSE, SRRMSE, and CC, respectively.