Work

A Critical Review of Adaptive Normalization Techniques for Cross-Session EEG Decoding on Embedded Platforms

Year: 2026
Type: preprint
Abstract: Cross-session non-stationarity remains a primary obstacle to the practical deployment of electroencephalography (EEG) decoding systems, as feature distributions drift substantially across recording sessions due to electrode repositioning, impedance fluctuations, and cognitive state variability. Adaptive normalization techniques have emerged as a computationally efficient countermeasure, aligning feature statistics in an online or calibration-light manner without requiring extensive labeled target data. However, the suitability of these methods for resource-constrained embedded platforms—where memory, latency, and energy budgets are severely limited—has not been systematically examined. This review critically evaluates adaptive normalization approaches for cross-session EEG decoding through an embedded-deployment lens. We categorize existing techniques into three families: exponential moving average-based feature normalization, adaptive instance normalization, and Riemannian or covariance-domain alignment. For each, we analyze computational complexity, memory footprint, sensitivity to hyperparameters, and compatibility with fixed-point or quantized inference. We further examine the trade-off between normalization efficacy and the overhead incurred by online statistic estimation, arguing that several widely adopted methods assume computational resources incompatible with microcontroller-class hardware. Our analysis identifies a gap between algorithmic sophistication and deployability, and we propose evaluation criteria that prioritize streaming operation, bounded memory, and robustness to distribution shift under quantization. We conclude that lightweight first-order normalization schemes currently offer the most favorable accuracy-to-cost ratio for embedded cross-session decoding, while geometry-aware and deep adaptation methods require architectural simplification before they can be considered viable on constrained platforms.
Author: Cheng Lee
Language: English
Open Access status: green
OpenAlex ID: W7236354047
Cites: 16
Cited by:
Related to: 10