The distribution regression problem encompasses many important statistics and machine learning tasks, and arises in a large range of applications. Among various existing approaches to tackle this problem, kernel methods have become a method of choice. Indeed, kernel distribution regression is both computationally favorable, and supported by a recent learning theory. This theory also tackles the two-stage sampling setting, where only samples from the input distributions are available. In this paper, we improve the learning theory of kernel distribution regression. We address kernels based on Hilbertian embeddings, that encompass most, if not all, of the existing approaches. We introduce the novel near-unbiased condition on the Hilbertian embeddings, that enables us to provide new error bounds on the effect of the two-stage sampling, thanks to a new analysis. We show that this near-unbiased condition holds for three important classes of kernels, based on optimal transport and mean embedding. As a consequence, we strictly improve the existing convergence rates for these kernels. Our setting and results are illustrated by numerical experiments.

We present the efficient implementations of probabilistic deterministic finite automaton learning methods available in FlexFringe. These implement well-known strategies for state-merging including several modifications to improve their performance in practice. We show experimentally that these algorithms obtain competitive results and significant improvements over a default implementation. We also demonstrate how to use FlexFringe to learn interpretable models from software logs and use these for anomaly detection. Although less interpretable, we show that learning smaller more convoluted models improves the performance of FlexFringe on anomaly detection, outperforming an existing solution based on neural nets.

We consider the problem of distribution-free learning for Boolean function classes in the PAC and agnostic models. Generalizing a beautiful work of Malach and Shalev-Shwartz (2022) that gave tight correlational SQ (CSQ) lower bounds for learning DNF formulas, we give new proofs that lower bounds on the threshold or approximate degree of any function class directly imply CSQ lower bounds for PAC or agnostic learning respectively. While such bounds implicitly follow by combining prior results by Feldman (2008, 2012) and Sherstov (2008, 2011), to our knowledge the precise statements we give had not appeared in this form before. Moreover, our proofs are simple and largely self-contained. These lower bounds match corresponding positive results using upper bounds on the threshold or approximate degree in the SQ model for PAC or agnostic learning, and in this sense these results show that the polynomial method is a universal, best-possible approach for distribution-free CSQ learning.

It is becoming increasingly important to understand the vulnerability of machine learning models to adversarial attacks. One of the fundamental problems in adversarial machine learning is to quantify how much training data is needed in the presence of evasion attacks, where data is corrupted at test time. In this thesis, we work with the exact-in-the-ball notion of robustness and study the feasibility of adversarially robust learning from the perspective of learning theory, considering sample complexity. We first explore the setting where the learner has access to random examples only, and show that distributional assumptions are essential. We then focus on learning problems with distributions on the input data that satisfy a Lipschitz condition and show that robustly learning monotone conjunctions has sample complexity at least exponential in the adversary's budget (the maximum number of bits it can perturb on each input). However, if the adversary is restricted to perturbing $O(\log n)$ bits, then one can robustly learn conjunctions and decision lists w.r.t. log-Lipschitz distributions. We then study learning models where the learner is given more power. We first consider local membership queries, where the learner can query the label of points near the training sample. We show that, under the uniform distribution, the exponential dependence on the adversary's budget to robustly learn conjunctions remains inevitable. We then introduce a local equivalence query oracle, which returns whether the hypothesis and target concept agree in a given region around a point in the training sample, and a counterexample if it exists. We show that if the query radius is equal to the adversary's budget, we can develop robust empirical risk minimization algorithms in the distribution-free setting. We give general query complexity upper and lower bounds, as well as for concrete concept classes.

Deep Neural Networks (DNNs) deployed to the real world are regularly subject to out-of-distribution (OoD) data, various types of noise, and shifting conceptual objectives. This paper proposes a framework for adapting to data distribution drift modeled by benchmark Continual Learning datasets. We develop and evaluate a method of Continual Learning that leverages uncertainty quantification from Bayesian Inference to mitigate catastrophic forgetting. We expand on previous approaches by removing the need for Monte Carlo sampling of the model weights to sample the predictive distribution. We optimize a closed-form Evidence Lower Bound (ELBO) objective approximating the predictive distribution by propagating the first two moments of a distribution, i.e. mean and covariance, through all network layers. Catastrophic forgetting is mitigated by using the closed-form ELBO to approximate the Minimum Description Length (MDL) Principle, inherently penalizing changes in the model likelihood by minimizing the KL Divergence between the variational posterior for the current task and the previous task's variational posterior acting as the prior. Leveraging the approximation of the MDL principle, we aim to initially learn a sparse variational posterior and then minimize additional model complexity learned for subsequent tasks. Our approach is evaluated for the task incremental learning scenario using density propagated versions of fully-connected and convolutional neural networks across multiple sequential benchmark datasets with varying task sequence lengths. Ultimately, this procedure produces a minimally complex network over a series of tasks mitigating catastrophic forgetting.

We characterize learnability for quantum measurement classes by establishing matching necessary and sufficient conditions for their PAC learnability, along with corresponding sample complexity bounds, in the setting where the learner is given access only to prepared quantum states. We first probe the results from previous works on this setting. We show that the empirical risk defined in previous works and matching the definition in the classical theory fails to satisfy the uniform convergence property enjoyed in the classical setting for some learnable classes. Moreover, we show that VC dimension generalization upper bounds in previous work are frequently infinite, even for finite-dimensional POVM classes. To surmount the failure of the standard ERM to satisfy uniform convergence, we define a new learning rule -- denoised ERM. We show this to be a universal learning rule for POVM and probabilistically observed concept classes, and the condition for it to satisfy uniform convergence is finite fat shattering dimension of the class. We give quantitative sample complexity upper and lower bounds for learnability in terms of finite fat-shattering dimension and a notion of approximate finite partitionability into approximately jointly measurable subsets, which allow for sample reuse. We then show that finite fat shattering dimension implies finite coverability by approximately jointly measurable subsets, leading to our matching conditions. We also show that every measurement class defined on a finite-dimensional Hilbert space is PAC learnable. We illustrate our results on several example POVM classes.

In 1984, Valiant [ 7 ] introduced the Probably Approximately Correct (PAC) learning framework for boolean function classes. Blumer et al. [ 2] extended this model in 1989 by introducing the VC dimension as a tool to characterize the learnability of PAC. The VC dimension was based on the work of Vapnik and Chervonenkis in 1971 [8 ], who introduced a tool called the growth function to characterize the shattering property. Researchers have since determined the VC dimension for specific classes, and efforts have been made to develop an algorithm that can calculate the VC dimension for any concept class. In 1991, Linial, Mansour, and Rivest [4] presented an algorithm for computing the VC dimension in the discrete setting, assuming that both the concept class and domain set were finite. However, no attempts had been made to design an algorithm that could compute the VC dimension in the general setting.Therefore, our work focuses on developing a method to approximately compute the VC dimension without constraints on the concept classes or their domain set. Our approach is based on our finding that the Empirical Risk Minimization (ERM) learning paradigm can be used as a new tool to characterize the shattering property of a concept class.

A classical result in online learning characterizes the optimal mistake bound achievable by deterministic learners using the Littlestone dimension (Littlestone '88). We prove an analogous result for randomized learners: we show that the optimal expected mistake bound in learning a class $\mathcal{H}$ equals its randomized Littlestone dimension, which is the largest $d$ for which there exists a tree shattered by $\mathcal{H}$ whose average depth is $2d$. We further study optimal mistake bounds in the agnostic case, as a function of the number of mistakes made by the best function in $\mathcal{H}$, denoted by $k$. We show that the optimal randomized mistake bound for learning a class with Littlestone dimension $d$ is $k + \Theta (\sqrt{k d} + d )$. This also implies an optimal deterministic mistake bound of $2k + \Theta(d) + O(\sqrt{k d})$, thus resolving an open question which was studied by Auer and Long ['99]. As an application of our theory, we revisit the classical problem of prediction using expert advice: about 30 years ago Cesa-Bianchi, Freund, Haussler, Helmbold, Schapire and Warmuth studied prediction using expert advice, provided that the best among the $n$ experts makes at most $k$ mistakes, and asked what are the optimal mistake bounds. Cesa-Bianchi, Freund, Helmbold, and Warmuth ['93, '96] provided a nearly optimal bound for deterministic learners, and left the randomized case as an open problem. We resolve this question by providing an optimal learning rule in the randomized case, and showing that its expected mistake bound equals half of the deterministic bound of Cesa-Bianchi et al. ['93,'96], up to negligible additive terms. In contrast with previous works by Abernethy, Langford, and Warmuth ['06], and by Br\^anzei and Peres ['19], our result applies to all pairs $n,k$.

To date, we know only a few handcrafted quantified Boolean formulas (QBFs) that are hard for central QBF resolution systems such as Q-Res and QU-Res, and only one specific QBF family to separate Q-Res and QU-Res. Here we provide a general method to construct hard formulas for Q-Res and QU-Res. The construction uses simple propositional formulas (e.g. minimally unsatisfiable formulas) in combination with easy QBF gadgets (Σb2 formulas without constant winning strategies). This leads to a host of new hard formulas, including new classes of hard random QBFs. We further present generic constructions for formulas separating Q-Res and QU-Res, and for separating Q-Res and LD-Q-Res.

Many of these analyses have focused on th e implications and uses of complexity-based algorithms defined by Blumer et a l. in two seminal papers [4, 5]. Their algorithms were defined such that they a chieved zero training error on a sample, and outputted a hypothesis whose complexity (VC dimension for continuous alphabets; description length for disc rete ones) was at most a polynomial in the target concept complexity, multiplied b y a sub-linear factor in the sam. These "Occam algorithms" are weak approx imations of the minimum-consistent-hypothesis problem [6]. In this paper, we focus on the continuous-alphabet Occam algorithms. In 1989, Blumer et al. [5] showed that if a concept was learnable by th eir Occam algorithm, then it was polynomially learnable; they left open the question of whether the converse of this theorem was true.