Machine Learning and AI have experienced unprecedented growth over the last ten years, especially on classical vision and language tasks. Expectations are high that the next decade of research will continue at the same pace, yielding transformative discoveries in nearly all areas of science. Complex machine perception tasks, however, still remain difficult. For example, tools for object recognition struggle with the dynamic nature and massive size of video data. What are the key foundational questions that need to be solved so that machine learning breaks through its current limitations? This institute (a partnership of The University of Texas at Austin, the University of Washington, Wichita State University, and Microsoft Research) will develop next-generation mathematical tools and algorithms with an eye towards major advances on core perceptual tasks. In the short term, principled heuristics will reduce the amount of trial and error that is prevalent in the current empirical framework. In the long term, new algorithms can reshape the landscape of machine learning with faster and more robust schemes for training and testing.
Singular goals of this institute include integrated plans for transferring knowledge that links foundational progress with use-inspired research spanning video, imaging, and navigation. These plans address outstanding issues in municipalities (e.g., traffic mitigation), the technology sector (e.g., robust video delivery), and healthcare (e.g., medical imaging). This institute addresses the intense industrial demand for expertise in machine learning by developing a new online master’s degree program in AI that caters to working professionals with complex schedules. Undergraduates will be brought to the forefront of machine learning research through a sequence of new research initiatives across several majors including computer science, mathematics, electrical engineering, and statistics. The institute also focuses on new internal strategies to dramatically increase the number of women in AI as well as new enrichment programs in machine learning for educators and students in high schools to broaden participation in AI for underserved communities.
The institute's research program identifies four key thrusts targeting major open problems in the foundations of machine learning: (i) Developing advanced algorithms for deep learning. The institute will create fast, provably efficient tools for training neural networks and searching parameter spaces. This includes formulating new analyses for gradient-based methods and applications to hyperparameter optimization and architecture search. (ii) Learning with dynamic data. Since datasets are constantly evolving, it is crucial to find algorithms and models that can incorporate changes at training and test time, including robustness to perturbations. A major emphasis will be on furthering a new field of efficient robust statistics, where the main objective is finding provably efficient algorithms for classical (often thought to be computationally intractable) statistical estimators. (iii) Exploiting Structure in Data. What characteristics of a dataset help with training and inference? Simple models involving sparsity often fail to capture modern training sets. This project will focus on more expressive ways to model the underlying structure in a distribution using deep networks as priors. Along these lines, the institute will investigate (both theoretically and experimentally) how structure can be used to address a fundamental mystery in supervised learning, namely why overparameterized networks generalize well when trained on real data. (iv) Optimizing real-world objectives. While black-box learning methods have improved by leaps and bounds in the recent past, it is difficult to use them in conjunction with complex loss functions that involve real-world constraints. For example, common vision tasks involve discriminating among different types of objects, but, depending on the downstream application, it may be more important to distinguish certain pairs of objects than others. For robot planning, mobility speed may be less important than adhering to certain safety constraints. This thrust will combine the current, optimization-based approach to machine learning with the expressive power of programming languages and formal methods. A related focus will be on imitation learning and interactive machine learning where objective functions can be adjusted based on real-world feedback from human users.
The institute's four research thrusts dovetail with three use-inspired research projects in video, imaging, and navigation. These projects encapsulate frontier challenges in machine perception and provide a wealth of benchmarks to evaluate theoretical work. With respect to video, the institute will work with industrial partners to redesign the whole video pipeline: from recognition and compression/decompression to training and model design. For imaging, the focus will be on new algorithms to solve ill-posed inverse problems where novel priors can compensate for the loss of observed information, especially in the context of medical diagnostics. The navigation project will consider the algorithmic challenge of autonomous and safe transportation in highly unstructured environments and will address policy issues in coordination with local governments. Altogether, these solutions will expand the scope of machine learning by moving beyond a black-box model of prediction.
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.
