Short-time multi-target SSVEP recognition using perceiver and large kernel convolution network

Steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) have garnered significant interest from researchers because of their high information transfer rate (ITR). Nevertheless, the efficacy of conventional recognition techniques is constrained by the length of the experimental data set and the number of stimulus targets. Although some studies have made progress in improving the recognition performance of SSVEP by employing deep learning algorithms, there is still significant room for performance enhancement. To address the challenges of SSVEP in short-time multi-target recognition, Our proposal involves developing a neural network based on deep learning for SSVEP that combines the Perceiver architecture and Large Kernel Depth-wise separable convolution, called PLKDNet. The network consists of an input convolutional layer targeting multiple electrode channels, a series of large kernel depth-wise separable convolutional layers, a cross-attention and self-attention mechanism that incorporates the Perceiver layer and an output layer. We evaluate the performance of PLKDNet on different datasets and compare it to two mainstream deep learning models under different training sample conditions. The comprehensive experimental results demonstrate that the PLKDNet model performs excellently in handling the short-time multi-target sample recognition task, and its classification accuracy outperforms that of the comparison algorithms, showing the potential to advance the recognition performance of SSVEP-based BCI systems.