Inverse Transfer Network with Frontier Point Restoration for EEG Transfer Classification

The transformation of data from the source domain into the target domain by minimizing their marginal and conditional distribution disparity is a common strategy in transfer learning. The distributions are measured using first-order, second-order, or higher-order data statistics. However, in brain-computer interface research, the scarcity of target-domain electroencephalograph (EEG) samples poses challenges for obtaining accurate statistics. To address this issue, we propose an inverse transfer scheme that relies solely on sufficient source-domain statistics to project target data back into the source domain. An end-to-end deep learning network comprising a feature extractor and a classifier is pretrained to establish a source feature space, whose distribution is described by the trained classifier. Then a replica of the trained feature-extractor is updated to map the target samples into this feature space. The feature extractor incorporates separable interference reduction and power spectrum extraction layers, enabling independent interference reduction transfer and spectrum alignment during the replica update. Due to the poor signal-to-noise ratio of EEG, interference reduction should be prioritized. To update the replica with the interference reduction transfer as the primary objective, a lightweight mapping block is employed to simplify the spectrum alignment. Furthermore, we proposed a method to suppress overfitting by increasing the intraclass distance. Extensive experiments verified that ITNet outperforms the state-of-the-art methods in EEG classification transfer tasks.