Foundations for Robust and Minimally-Supervised Learning Systems

点击翻译按钮获取中文摘要

Background. Machine learning, a technology that develops software through learning from historical experience and inferencing the properties of new data by its generalization ability, has been largely driven by the availability of big data. However, labeling of big data by human annotators is costly, and machine learning generalization ability is often susceptible to adversarial attacks and out-of-distribution data. For example, a tiny adversarial patch attached to a stop sign could completely fool autopilot systems, even though the systems were trained with as many as 3-billion-mile real data collected over 6 years, quoted from Tesla's 2020 report. Objectives. The long-term goal of the proposed research program is to advance machine learning from the robustness and self-supervision perspectives, by developing generic, principled, and scalable algorithms with provable guarantees and practical applications. Therefore, the objectives include: 1) developing new robust models against unrestricted adversarial examples and out-of-distribution data and connecting robustness with fairness, privacy and interpretability towards trustworthy machine learning; 2) closing the performance gap between supervised and self-supervised learning; and 3) applying the proposed robust and self-supervised framework to the real-world tasks, including 3D reconstruction, textual analysis, and Deepfake detection. Impacts and Reward. The developed robust method is expected to serve as the first baseline to encourage the development of new defenses against a broader class of adversarial attacks. The technology will potentially accelerate the research and development of autonomous cars, making AI safe, reliable, transparent, and trustworthy in Canada. Moreover, this research will significantly reduce the cost of data labeling for high-tech companies in Canada. The proposed research program will improve the state-of-the-art of real-time 3D perception for robotics navigation, AR/VR for Metaverse, textual analysis in the presence of typos, and fake video detection on YouTube and Twitter. Students will learn the necessary skills to carry out a research plan, scientific writing, critical thinking, and coding. By the end of their program, the students will become machine learning experts who are well-positioned for industrial research scientist positions or academic research positions in top universities.