Abstract
Human computer interaction (HCI) involves a multidisciplinary fusion oftechnologies, through which the control of external devices could be achievedby monitoring physiological status of users. However, physiological biosignalsoften vary across users and recording sessions due to unstable physical/mentalconditions and task-irrelevant activities. To deal with this challenge, wepropose a method of adversarial feature encoding with the concept of a RatelessAutoencoder (RAE), in order to exploit disentangled, nuisance-robust, anduniversal representations. We achieve a good trade-off between user-specificand task-relevant features by making use of the stochastic disentanglement ofthe latent representations by adopting additional adversarial networks. Theproposed model is applicable to a wider range of unknown users and tasks aswell as different classifiers. Results on cross-subject transfer evaluationsshow the advantages of the proposed framework, with up to an 11.6% improvementin the average subject-transfer classification accuracy.