A research team from Xuzhou Medical University has introduced FreqW-LiteNet, a lightweight neural network designed for cross-session electroencephalography emotion recognition that operates using approximately 23,000 parameters. Published in Frontiers in Neuroscience, the study demonstrates that the compact architecture achieves stable performance across separate recording sessions by utilizing frequency-band differentiated weighting, offering a viable solution that balances performance and efficiency for neural signal decoding in resource-constrained scenarios.
Non-invasive electroencephalography holds promising potential for brain-computer interfaces and neural engineering applications. Existing analytical models often demand heavy computational resources, creating major engineering bottlenecks for practical on-device deployment.
To address these persistent challenges, researchers Rui Fan, Houru Yan, ZhengXiang Wan, and Junfeng Hu began by conducting progressive frequency-band ablation experiments. Their experimental analysis first verified the effectiveness of full-band features. The ablation testing also revealed distinct complementary patterns between high-frequency and low-frequency features, providing an empirical foundation for applying differentiated weighting across frequency bands.
Building directly upon these findings, the team developed FreqW-LiteNet around a single-backbone fully connected architecture. Rather than implementing computationally expensive mechanisms such as multi-source domain adaptation or adversarial training, the model relies on frequency-band differentiated weighting to extract informative emotional cues while minimizing structural complexity.
To ensure practical efficiency alongside structural simplicity, the researchers implemented a two-stage training strategy combined with high-intensity regularization. This architectural and training synergy allows FreqW-LiteNet to achieve efficient training directly on standard central processing units, removing the necessity for specialized or power-intensive computing hardware.
Systematic validation was conducted on three public benchmarks: the SEED, SEED-IV, and SEED-V datasets. All experimental evaluations were strictly restricted to these SEED-family datasets and carried out under cross-session protocols within a transductive evaluation setting, specifically testing the system’s capacity to handle distribution shifts between independent recording sessions.
Experimental results demonstrated that FreqW-LiteNet achieved competitive emotion recognition accuracy on the SEED and SEED-IV datasets compared with state-of-the-art domain adaptation methods and general deep learning models. The authors noted that the model is extremely compact, containing approximately 23,000 parameters, which considerably reduces computational overhead while delivering stable cross-session emotion recognition for wearable applications.
Subject of Research: Cross-session EEG emotion recognition using a lightweight neural network
Article Title: A lightweight cross-session EEG emotion recognition network based on high- and low-frequency differential weighting
Article References: Fan, R., Yan, H., Wan, Z., & Hu, J. (2026). A lightweight cross-session EEG emotion recognition network based on high- and low-frequency differential weighting. Frontiers in Neuroscience, 20, Article 1992372. https://doi.org/10.3389/fnins.2026.1992372
Image Credits: AI Generated
DOI: 10.3389/fnins.2026.1992372
Keywords: electroencephalography, EEG, emotion recognition, cross-session, FreqW-LiteNet, frequency-band weighting, lightweight network, brain-computer interfaces, neural engineering, wearable devices, SEED dataset, SEED-IV, SEED-V, two-stage training, regularization, domain adaptation, transductive evaluation
News Source: Cassandra Pierce. (November 19, 2026). Lightweight Neural Network Decodes EEG Emotion Across Sessions With High Efficiency. Scienmag.



