CAREER: Learning at the Edge: an Extreme Value Theory for Visual Recognition

点击翻译按钮获取中文摘要

The brain has the remarkable ability to rapidly and accurately extract meaning from a flood of complex and ever-changing sensory information. A key question is how neuronal systems encode relevant information about the external world, especially with respect to perceptual tasks such as object recognition and categorization. Unlocking this secret would likely change the typical way in which we approach machine learning for computer vision --- a key area of artificial intelligence. This project seeks to answer this question using methods and procedures from psychology and statistics, leading to new AI capabilities that can be transitioned to commercial and government applications where the processing of visual information is a concern. The recognition model produced could have tremendous impact across a number of fields including computer vision, machine learning, neuroscience, psychology, and cognitive science. In addition, the PI will train students at the undergraduate and graduate levels, and organize new workshops relating computer science to human vision.The project’s first technical objective is to work towards a potentially transformative Extreme Value Theory (EVT) for visual recognition. This new theoretical framework will provide solid grounding for vision scientists and AI engineers working on experiments that model the human visual system. It will facilitate both new theoretical analyses of decision making in a visual context, as well as new classification algorithms that are more biologically-consistent in operation. The second technical objective is an experimental assessment of the EVT recognition model. The significant difference in predictions made by EVT and more conventional modeling strategies that rely on central tendency assumptions allows us to formulate testable hypotheses that support psychophysical studies to understand the role of extrema in recognition. The high-level design is three experiments in two different regimes, with humans as the focus of study. This study aims to uncover the principles of decision making that underpin object recognition. The third technical objective is to use the developed theory and the experimental results to create a new class of biologically-consistent machine learning algorithms for decision making that are a measurable advance beyond the state-of-the-art. This includes generative learning algorithms to model the distributions for specific class representations, and discriminative learning algorithms to find the boundaries between classes. This effort will develop probabilistic generative EVT mixture models, as well as probabilistic discriminative one-class, binary, and multi-class EVT classifiers.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.