Adaptivity in Online and Statistical Learning

Many modern machine learning algorithms, though successful, are still based on heuristics. In a typical application, such heuristics may manifest in the choice of a specific Neural Network structure, its number of parameters, or the learning rate during training. Relying on these heuristics is not ideal from a computational perspective (often involving multiple runs of the algorithm), and can also lead to over-fitting in some cases. This motivates the following question: for which machine learning tasks/settings do there exist efficient algorithms that automatically adapt to the best parameters? Characterizing the settings where this is the case and designing corresponding (parameter-free) algorithms within the online learning framework constitutes one of this thesis' primary goals. Towards this end, we develop algorithms for constrained and unconstrained online convex optimization that can automatically adapt to various parameters of interest such as the Lipschitz constant, the curvature of the sequence of losses, and the norm of the comparator. We also derive new performance lower-bounds characterizing the limits of adaptivity for algorithms in these settings. Part of systematizing the choice of machine learning methods also involves having ``certificates'' for the performance of algorithms. In the statistical learning setting, this translates to having (tight) generalization bounds. Adaptivity can manifest here through data-dependent bounds that become small whenever the problem is ``easy''. In this thesis, we provide such data-dependent bounds for the expected loss (the standard risk measure) and other risk measures. We also explore how such bounds can be used in the context of risk-monotonicity