Smoothed Analysis for Learning Concepts with Low Intrinsic Dimension
Explore a conference talk on smoothed analysis for learning concepts with low intrinsic dimension. Delve into a novel smoothed analysis framework that addresses hardness results in the agnostic learning model. Discover how this approach enables learning results for concepts dependent on low-dimensional subspaces and with bounded Gaussian surface area. Examine applications to functions of halfspaces and low-dimensional convex sets, previously only learnable under structured distributions. Uncover new insights for traditional non-smoothed frameworks, including an improved algorithm for agnostically learning intersections of halfspaces. Gain valuable knowledge on computational vs statistical gaps in learning and optimization from this presentation at IPAM's EnCORE Workshop.