Self-Conscience/Physical Memory is a brain-controlled robotic sculpture in the University of Houston’s Noninvasive Brain Machine Interface Laboratory. Motorized and lluminated acrylic ceiling tiles shift the architecture of the space itself in response to EEG data. The height of the panels is driven by alpha power suppression in the central cortical areas, and the tiles’ color shifts with alpha power changes in the occipital and frontal lobes. The EEG data can be input in real-time by a single participant, or, in the absence of user input, the work also serves as a playback device for archival EEG recordings, a physical manifestation of a past experience, of a moment in someone’s life, a person both absent and present in that new moment.

Abstract: The necessity for expedient documentation of the actions performed by subjects under recorded observation is necessary for many fields of research to effect the creation of a reference database for correlation of actions and stimuli. This research in particular sought to develop a program capable of documenting the actions a research subject was performing while wearing a cap designed to capture electroencephalography (EEG) signals. The subject in question was under observation for several months in total, which resulted in hundreds of hours of videos requiring documentation to understand how the subject’s activities would appear when represented in EEG. Documenting this amount of video by hand would require weeks of work by a team of volunteers typically. With the creation of software which takes advantage of the latest techniques in machine vision for object detection and recognition, this research resulted in the development of a rudimentary system to accomplish the same goal of video annotation automatically without any supervision from a human. The program was developed and then used to segment 11 sample videos for practical use-case quantification in the identification of classification accuracy across five classes of actions. The program was able to classify 88% of frames correctly across these sample videos.

Abstract: Classifying EEG signals using state of the art deep learning techniques still leaves much to be desired. EEG is known for issues related to statistical noise as well as variance between subjects and even between sessions. The seminal research paper written by Schirrmeister et al., “Deep Learning With Convolutional Neural Networks for EEG Decoding and Visualization” described a deep learning architecture that was shown to be capable of classifying EEG with an accuracy of 92.4%. The EEG in question was EEG captured during the movement of hands and feet of research subjects. The architecture developed by Schirrmeitster et al. was also used to extract feature meaningfulness. Through perturbation of data and analysis of the subsequent change in classification accuracy, the researchers were capable of discovering the EEG channels and neuronal activation frequency bands most necessary in each class. This research was replicated with the intent of classifying actions related to the art production process while preserving the architecture developed in the original paper.

Abstract: The Capsule Network neural network architecture was developed recently by Hinton et al. as an attempt to imitate brain functionality and produce a more effective neural network architecture. The Capsule Network has been shown to be effective at classification of the MNIST hand written digit data set with 99.7% Accuracy while requiring less than 30 epochs to be trained to do so. It has been theorized that the capsule network will be resistant to issues effecting classification accuracy due to noise present in the input data which, if true could be a massive benefit to EEG classification. This research documents the use of The Capsule Network for classifying EEG data alongside more understood techniques.

Abstract: The aim of this project is to utilize machine learning methods developed for genomic sequencing to more effectively analyze binary executables for feature in code which will reliably indicate security vulnerabilities. First, by translating the binary to a DNA base analogue then, uploading this information to the online genome sequencing algorithm mVista, and finally processing these results the researchers were able to discern a set of features. Then with this information were able to identify the levels of occurence of the features in a set of 10 safe sets of code and a set of 10 unsafe pieces of code. Then through data analytics, the researchers were able to discern the unsafe from the safe code with 92% accuracy.