ECE Associate Professor Waheed Bajwa has been awarded a grant from the Army Research Office (ARO) for the project titled "Statistical learning for the modern datasets: Generalization bounds and near-optimal learning algorithms." The 3-year, $360,000 award will advance the state-of-the-art in statistical learning theory and lead to computationally efficient algorithms for machine learning. While the statistical learning framework has long played a central role in advancing our understanding of machine learning systems, there is an interest in looking afresh at the questions of generalization error bounds, fundamental limits, and near-optimal algorithms in the face of modern datasets that increasingly represent a ‘zoo’. Indeed, the classical statistical learning works typically focused on centralized datasets that often had Euclidean geometry. In contrast, many of today’s and tomorrow’s applications of machine learning involve non-Euclidean datasets that are non-centralized, with data often streaming at very high rates, some of which might be compromised due to either gross errors or actions of adversarial entities. Such modern datasets necessitate development of fundamentally new analytical tools and algorithmic techniques for statistical learning-based study of machine learning systems. It is in this regard that this project leverages tools from stochastic approximation, (centralized and distributed) optimization theory, concentration-of-measure literature, information theory, robust statistics, and tensor algebra to derive generalization error bounds, fundamental limits on sample complexity, and near-optimal learning algorithms for machine learning from modern datasets. The outcomes of this project are expected to not only advance the state-of-the-art in statistical learning theory, but they are also expected to lead to computationally efficient algorithms for machine learning that can be deployed in practical settings with the smallest number of training samples.
Congratulations, Waheed!