A deep learning approach to training a brain activity-based trial-by-trial classifier for rapid serial visual presentation imagery
- Katherine Porterfield, Pacific Northwest National Laboratory, Seattle, Washington, United States
- Leslie Blaha, Pacific Northwest National Laboratory, Richland, Washington, United States
- Jonathan Suter, Pacific Northwest National Laboratory, Richland, Washington, United States
- Johnathan Cree PH.D, Pacific Northwest National Laboratory, Richland, Washington, United States
- Gianluca Longoni, Pacific Northwest National Laboratory, Seattle, Washington, United States
- Jesse Johns, Pacific Northwest National Laboratory, Richland, Washington, United States
- Gerges Dib, Pacific Northwest National Laboratory, Richland, Washington, United States
- Kayla Duskin, Pacific Northwest National Laboratory, Richland, Washington, United States
AbstractImage classification aided by brain activity measured during rapid serial visual presentation (RSVP) shows promise to aid human viewers to quickly triage large volumes of images with support of an EEG technology. Fast perceptual responses are parsed with a brain-activity classifier operating on EEG signals to select an image subset containing visual information similar to the viewer’s target. However, current processes for training brain activity classifiers are experimentally and computationally expensive. We propose a deep learning model that classifies images based off of brain-activity. Using the satellite visual images and EEG data provided from Bigdely-Shamlo et al. (2007), we compare different machine learning (Support Vector Machines) and deep learning (Convolutional Neural Networks and Recurrent Neural Networks) approaches along with different data manipulation styles for classifying the satellite images. This initial report summarizes the efforts to establish benchmarks for deep learning, exploring the potential to streamline and improve brain-activity based classification.