Seminar | February 25 | 11:10 a.m.-12:30 p.m. | 489 Minor Hall
Dylan Paiton, Helen Wills Neuroscience Institute; Elise Harb, UC Berkeley School of Optometry
"Neural Response Geometry Explains Adversarial Directions"
Lateral connections abound in biological neural networks, but are absent from the feedforward architectures popular in artificial neural networks. Artificial neural networks are also vulnerable to adversarial examples: small perturbations to the input can produce large changes in the output. In this work, we demonstrate that a network with lateral connectivity in its hidden layer requires significantly larger input perturbations to achieve a change in its output, as compared to a feedforward network. Furthermore, we present evidence that neurons with population nonlinearities (via lateral connections) have a higher degree of selectivity to input directions that lie along the data manifold, as revealed by the curved geometry of their response fields. We hypothesize that this selectivity forces adversarial attacks to be more semantically similar to the attack target, resulting in data aligned adversarial attacks.
Elise Harb, OD
"Human Behaviors and Myopia: A Novel Approach"
The rapid rise in the prevalence of myopia worldwide is a significant public health concern and implicates environmental rather than genetic factors in its development. Several environmental factors have been implicated in the development or progression of myopia; however, the mechanism is not clear and is likely multi-factorial. Significant attention has been paid to the possible risk effects of near work and, more recently, the protective effects of outdoor activity on the development and/or progression of myopia, but what aspects of these activities are responsible for the observed effects remain to be determined. This may be in part due to the traditional use of subjective questionnaires as a measurement tool of habitual behavior. We aim to take a novel approach by using more objective techniques, including wearable technologies, to better understand the dynamic behaviors of children and their potential role in myopia development and/or progression.