Knowledge-driven feature component interpretable network for motor imagery classification

Abstract

Objective. The end-to-end convolutional neural network (CNN) has achieved great success in motor imagery (MI) classification without a manual feature design. However, all the existing deep network solutions are purely datadriven and lack interpretability, which makes it impossible to discover insightful knowledge from the learned features, not to mention to design specific network structures. The heavy computational cost of CNN also makes it challenging for real-time application along with high classification performance. Approach. To address these problems, a novel knowledge-driven feature component interpretable network (KFCNet) is proposed, which combines spatial and temporal convolution in analogy to independent component analysis and a power spectrum pipeline. Prior frequency band knowledge of sensory-motor rhythms has been formulated as band-pass linear-phase digital finite impulse response filters to initialize the temporal convolution kernels to enable the knowledge-driven mechanism. To avoid signal distortion and achieve a linear phase and unimodality of filters, a symmetry loss is proposed, which is used in combination with the cross-entropy classification loss for training. Besides the general prior knowledge, subject-specific time-frequency property of event-related desynchronization and synchronization has been employed to construct and initialize the network with significantly fewer parameters. Main results. Comparison of experiments on two public datasets has been performed. Interpretable feature components could be observed in the trained model. The physically meaningful observation could efficiently assist the design of the network structure. Excellent classification performance on MI has been obtained. Significance. The performance of KFCNet is comparable to the state-of-the-art methods but with much fewer parameters and makes real-time applications possible.