R&D Division, Brain Corporation, San Diego, CA
3-4:30 p.m. Monday, April 18, 2016
Real-world immersion of artificial neural networks through robotics
Much like real biological organisms, autonomous robots will need to perceive the world around them. Despite recent advances in machine learning for vision, robots today remain confined to tightly controlled physical spaces — or often require the intervention of a human. In this talk I outline an R&D methodology that we are using to design new kinds of machine learning algorithms aimed at this problem. The methodology involves two parts: (1) new, appropriate benchmarks and metrics, and (2) a software infrastructure called BrainOS that enables us to rapidly develop and seamlessly evaluate algorithms across datasets, photorealistic virtual worlds, and real robots. Because real world perception depends heavily on time and context, we designed our neural network architecture around concepts of recurrent processing, prediction (e.g., Elman, 1990; Mikolov et al., 2015) and prediction errors (Clark, 2013). In contrast, deep convolutional neural nets (“deep learning”) ignore time and context. As a first quantifiable task for our algorithm, we chose visual object tracking. Preliminary results show our network can visually track objects with performance on par with or better than state-of-the-art, expert-crafted visual tracking algorithms. These results hold under challenging perceptual conditions like those faced by robots moving around in the world (e.g., variations in illumination, shadow, motion blur). These findings give us cause for optimism that new generations of neural network architectures will be able to deal with the “grittiness” of real world perception.
Patryk Laurent, a scientist at Brain Corporation since 2012, specializes in the cognitive science of learning and memory. He holds a bachelor’s degree in the Cognitive Sciences from the University of Virginia, and earned his doctorate in Neuroscience at the University of Pittsburgh, with affiliations at the Center for the Neural Basis of Cognition (CNBC) and the Learning Research and Development Center (LRDC). Then, Dr. Laurent spent three years as a postdoctoral researcher at the Johns Hopkins University, investigating the influence of reward on visual attention and perception. He has published over a dozen peer-reviewed articles in computational and human experimental research. In industrial research at Brain Corporation, Patryk investigates and applies principles of learning systems to robotics. He is part of a team that studies how learning and design principles inspired by the brain allow computational and robotics systems to successfully immerse themselves in, and interact with, the real world.