RI: Medium: Deep Annotation: Measuring Human Vision to Improve Machine Vision

点击翻译按钮获取中文摘要

Machine learning is the science of designing computational systems that can learn from data, much as humans do. However, while many machine learning approaches rely on humans to provide labels for training examples that are used for learning, human-provided labels represent just a tiny fraction of the information that can be gleaned from humans. This project brings together a multidisciplinary team with expertise spanning computer science, neuroscience and psychology to pioneer a new paradigm in machine learning that seeks to better mimic human performance by incorporating new kinds of information about human behavior. Specifically, this project brings the disciplines of psychophysics and psychometrics, which seek to quantitatively describe human performance -- patterns of errors, reaction times, and variations across populations of humans -- together with machine learning to develop systems that learn both from human successes and failures, to generate artificial systems that perform better and generalize better to data outside of their training sets. The project team has already shown initial proof of concept in applying these ideas to the problem of face detection in difficult, cluttered real-world images. During the project period, the team will greatly expand these ideas, developing new applications (including face and object recognition tasks), a broader range of machine learning settings (including regression and feature selection), and methods for incorporating new kinds of data (such as fMRI brain scans) for guiding machine learning algorithms. This research represents a new direction in machine learning research, which increasingly has important and broad impact in our modern, data-driven world. In addition, it is anticipated that the theoretical gains in machine learning derived from this work will feed back into psychology, enabling rapid screening of candidate hypotheses about how the brain works by artificial systems which can then be tested on humans using an advanced crowdsourcing platform for quantifying human behavior.