Decompose Time and Frequency Dependencies: Multivariate Time Series Physiological Signal Emotion Recognition
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Physiological Signal, Emotion Recognition, Time Series, Representation Learning, Affective Computing
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In this study, we proposed a transformer based end-to-end solution to capture the relationship between the physiological signals and affective changes. We first convert the physiological signal emotion recognition prediction task to a sequence-to-sequence multivariate time series prediction task. We utilize the state-of-the-art (SOTA) self-attention mechanism to decompose the physiological signals into separate frequency domain and time domain representations, and capture the channel dependencies via Two-Stage Attention (TSA). Meanwhile, we implement the multitask learning framework to better predict the valence and arousal affective states individually. We evaluate our system on the Continuously Annotated Signals of Emotion (CASE) dataset used in the Emotion Physiology and Experience Collaboration (EPiC) challenge, and our proposed system outperform all the challenge participants in all four test scenarios.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8530
Loading