“…Interestingly, the weights in the matrix of random projection do not need to be stored (i.e., can be rematerialized by a random function on the fly), or can be realized by emerging memristor [16], [17], [18], [19], [20] and optical [21] devices. A readout function layer can then effectively analyze the projected features for various classification tasks, e.g., in EEG [22], [23], electrocardiography (ECG) signals [24], [25], and electrocorticography (ECoG) [26]. On the other hand, for the CNN-based approaches in MI-BCIs, quantization methods to 8-bit fixed-point weights and activations are developed [27], but having a CNN model with full, or partial, binary weights is still missing in MI-BCIs.…”