Evolving neural networks through augmenting topologies
Evolutionary Computation
Picbreeder: evolving pictures collaboratively online
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems
Using physiological signals to evolve art
EuroGP'06 Proceedings of the 2006 international conference on Applications of Evolutionary Computing
Hi-index | 0.00 |
We propose a brain-computer interface (BCI) system for evolving images in real-time based on subject feedback derived from electroencephalography (EEG). The goal of this system is to produce a picture best resembling a subject's `imagined' image. This system evolves images using Compositional Pattern Producing Networks (CPPNs) via the NeuroEvolution of Augmenting Topologies (NEAT) genetic algorithm. Fitness values for NEAT-based evolution are derived from a real-time EEG classifier as images are presented using rapid serial visual presentation (RSVP). Here, we report the design and performance, for a pilot training session, of a BCI system for real-time single-trial binary classification of viewed images based on participant-specific brain response signatures present in 128-channel EEG data. Selected training-session image clips created by the image evolution algorithm were presented in 2-s bursts at 8/s. The subject indicated by subsequent button press whether or not each burst included an image resembling two eyes. Approximately half the bursts included such an image. Independent component analysis (ICA) was used to extract a set of maximally independent EEG source time-courses and their 100 minimally-redundant low-dimensional informative features in the time and time-frequency amplitude domains from the (94%) bursts followed by correct manual responses. To estimate the likelihood that the post-image EEG contained EEG `flickers' of target recognition, we applied two Fisher discriminant classifiers to the time and/or time-frequency features. The area under the receiver operating characteristic (ROC) curve by tenfold cross-validation was 0.96 using time-domain features, 0.97 using time-frequency domain features, and 0.98 using both domain features.