Bracha Laufer - Leveraging Structures in Complex Spaces for Robust, Reliable and Efficient Learning

With the great promise of the data-science revolution, a major open question is whether the underlying models are efficient and trustworthy, especially when deployed in complex real-world settings. In this talk I will show how these challenges can be approached from a data-driven perspective, relying on the geometry and the statistical patterns hidden in the data to establish strong notions of robustness and reliability. In the first part I will present several geometry-based methods for extracting intrinsic representations from noisy multi-path and multi-source acoustic measurements. These representations simplify intricate physical acoustic models and provide the key for deriving robust and efficient audio processing tools. Specifically, I will present methodologies for structured multi-view fusion, learning dynamics over manifolds, hybrid data-driven and classical inference, and simplex representations for source separation. In the second part I will focus on statistical tools for uncertainty quantification and control of complex machine learning models. I will present a new framework for selecting model configurations that provably satisfy multiple explicit and simultaneous statistical guarantees, while also optimizing any number of additional, unconstrained objectives. The method combines multi-objective optimization and statistical testing, creating a framework that is both computationally and statistically efficient, even when the configuration space is combinatorially large. I will show how this method can reliably and adaptively solve the timely problem of long inference times of large-scale Transformer models for text classification under various user-defined performance specifications.