Deep learning for rapid serial visual presentation event from electroencephalography signal

The goal of bio-inspired machine research is to create and apply a new-generation interaction between human and machine for our life. It allows machine to be directly triggered by users' simultaneous responses. Brain computer interface (BCI), which allows human subjects to communicate with or control an external device with their brain signals, belongs to one type of such research. There are many paradigms for BCI systems, among which one paradigm named rapid serial visual presentation (RSVP) tasks. RSVP aims at collecting personal EEG when tested subjects are asked to identify a target image from a continuous burst of image clips presented at a high rate. Subjects are required to search for target images from a large collection of undesirable ones, and the target image can be predefined or decided by certain rules. Our goal for RSVP tasks is to use machine learning, especially deep learning algorithms to predict whether subjects have seen a target or not.

We performed a comprehensive investigation on deep learning algorithm based classification predicting the target vs. non-target EEG epochs. The testing scenarios includes both time-locked RSVP tasks and non-time locked RSVP tasks with 6 different EEG experiments. The investigation of deep learning algorithm includes deep stacking network (DSN), deep neural network (DNN) and deep convolutional network (CNN). We also proposed the feature visualization methods for DSN and investigated the deconvolutional network as the visualization technique for CNN. The deep learning transferability is also investigated by our proposed DSN transfer learning and CNN transfer learning model on RSVP data. In addition, calibration sample size for classification in BCI systems have also been investigated and new feature combinations that will provide robust improvement in RSVP classification accuracy have been tested. In sum, we have studied DL solutions to classify BCI, especially RSVP tasks and provide methods dealing with experiment and subject dynamics for BCI tasks.