Ensemble diverse hypotheses and knowledge distillation for unsupervised cross-subject adaptation

Abstract

Human intent prediction (HIP) and human activity recognition (HAR) are important for human–robot interactions. However, human–robot interface signals are user-dependent. A classifier trained on labeled source subjects performs poorly on unlabeled target subjects. Besides, previous methods used a single learner, which may only learn a subset of features and degrade their performance on target subjects. Last, HIP and HAR require real-time computing on edge devices whose computational capabilities limit the model size. To address these issues, this paper designs an ensemble diverse hypotheses (EDH) and knowledge distillation (EDHKD) method. EDH mitigates the cross-subject divergence by training feature generators to minimize the upper bound of the classification discrepancy among multiple classifiers. EDH also maximizes the discrepancy among multiple feature generators to learn diverse and complete features. After training EDH, a lightweight student network (EDHKD) distills the knowledge from EDH to a single feature generator and classifier to significantly decrease the model size but remain accurate. The performance of EDHKD is theoretically demonstrated and experimentally validated. Results show that EDH can learn diverse features and adapt well to unknown target subjects. With only soft labels provided by EDH, the student network (EDHKD) can inherit the knowledge learned by EDH and classify unlabeled target data of a 2D moon dataset and two human locomotion datasets with the accuracy at 96.9%, 94.4%, and 97.4%, respectively, in no longer than 1 millisecond. Compared to the benchmark method, EDHKD lifts the target-domain classification accuracy by 1.3% and 7.1% in the two human locomotion datasets. EDHKD also stabilizes learning curves. Therefore, EDHKD significantly increases the generalization ability and efficiency of the HIP and HAR.