An adaptive cross-class transfer learning framework with two-level alignment

Abstract

Because of insufficient data, cross-class transfer learning has potential prospects in motor imagery electroencephalogram (MI-EEG) based rehabilitation engineering, and it has not been effectively addressed to reduce cross-class variability between source and target domains and find the best class-to-class transitive correspondence (CCTC). In this paper, we propose an adaptive cross-class transfer learning (ACTL) framework with two-level alignment (TLA) for MI decoding. By using the fast partitioning around medoids method, the central samples of each class are acquired from the source domain and target domain respectively. Then, they are applied to construct the central alignment matrix to realize the 1st-level domain alignment for any possible CCTC case. The 2nd-level subject alignment is performed by Euclidean alignment for each class of aligned source domain. Furthermore, the central samples of the target domain together with the two-level aligned source domain are mapped into tangent space, the resulting features are fed to a teacher convolutional neural network (TCNN), it and the transferred student CNN (SCNN) are trained successively, and their parameters of distillation loss are optimized automatically by a scaling-based grid search method. Experiments are conducted on a public MI-EEG dataset with multiple subjects and MI-tasks, the proposed framework can find the optimal SCNN associated with the best CCTC, which achieves a statistically significant average classification accuracy of 86.37%. The results suggest that TLA is helpful for increasing the distribution similarity between different domains, and the knowledge distillation embedded in ACTL framework greatly simplifies the SCNN and outperforms the complicated TCNN.