The main goal of this article is to convince you, the reader, that supervised learning in the Probably Approximately Correct (PAC) model is closely related to -- of all things -- bipartite matching! En-route from PAC learning to bipartite matching, I will overview a particular transductive model of learning, and associated one-inclusion graphs, which can be viewed as a generalization of some of the hat puzzles that are popular in recreational mathematics. Whereas this transductive model is far from new, it has recently seen a resurgence of interest as a tool for tackling deep questions in learning theory. A secondary purpose of this article could be as a (biased) tutorial on the connections between the PAC and transductive models of learning.
We study ``selective'' or ``conditional'' classification problems under an agnostic setting. Classification tasks commonly focus on modeling the relationship between features and categories that captures the vast majority of data. In contrast to common machine learning frameworks, conditional classification intends to model such relationships only on a subset of the data defined by some selection rule. Most work on conditional classification either solves the problem in a realizable setting or does not guarantee the error is bounded compared to an optimal solution. In this work, we consider selective/conditional classification by sparse linear classifiers for subsets defined by halfspaces, and give both positive as well as negative results for Gaussian feature distributions. On the positive side, we present the first PAC-learning algorithm for homogeneous halfspace selectors with error guarantee $\bigO*{\sqrt{\mathrm{opt}}}$, where $\mathrm{opt}$ is the smallest conditional classification error over the given class of classifiers and homogeneous halfspaces. On the negative side, we find that, under cryptographic assumptions, approximating the conditional classification loss within a small additive error is computationally hard even under Gaussian distribution. We prove that approximating conditional classification is at least as hard as approximating agnostic classification in both additive and multiplicative form.
The notion of rank of a Boolean function has been a cornerstone in the theory of PAC learning, enabling quasipolynomial-time learning algorithms for polynomial-size decision trees. We present a novel characterization of rank, grounded in the well-known Transformer architecture. We show that the rank of a function $f$ corresponds to the minimum number of Chain of Thought (CoT) steps required by a single-layer transformer decoder with hard attention to compute $f$. Based on this characterization we establish tight bounds on the number of CoT steps required for specific problems, showing that $\ell$-fold function composition necessitates exactly $\ell$ CoT steps. Furthermore, we analyze the problem of identifying the position of the $k$-th occurrence of 1 in a Boolean sequence, proving that it requires $k$ CoT steps.
The concepts of linkage, building blocks, and problem decomposition have long existed in the genetic algorithm (GA) field and have guided the development of model-based GAs for decades. However, their definitions are usually vague, making it difficult to develop theoretical support. This paper provides an algorithm-independent definition to describe the concept of linkage. With this definition, the paper proves that any problems with a bounded degree of linkage are decomposable and that proper problem decomposition is possible via linkage learning. The way of decomposition given in this paper also offers a new perspective on nearly decomposable problems with bounded difficulty and building blocks from the theoretical aspect. Finally, this paper relates problem decomposition to PAC learning and proves that the global optima of these problems and the minimum decomposition blocks are PAC learnable under certain conditions.
We study the problem of PAC learning $\gamma$-margin halfspaces with Massart noise. We propose a simple proper learning algorithm, the Perspectron, that has sample complexity $\widetilde{O}((\epsilon\gamma)^{-2})$ and achieves classification error at most $\eta+\epsilon$ where $\eta$ is the Massart noise rate. Prior works [DGT19,CKMY20] came with worse sample complexity guarantees (in both $\epsilon$ and $\gamma$) or could only handle random classification noise [DDK+23,KIT+23] -- a much milder noise assumption. We also show that our results extend to the more challenging setting of learning generalized linear models with a known link function under Massart noise, achieving a similar sample complexity to the halfspace case. This significantly improves upon the prior state-of-the-art in this setting due to [CKMY20], who introduced this model.
We study the problem of PAC learning $\gamma$-margin halfspaces in the presence of Massart noise. Without computational considerations, the sample complexity of this learning problem is known to be $\widetilde{\Theta}(1/(\gamma^2 \epsilon))$. Prior computationally efficient algorithms for the problem incur sample complexity $\tilde{O}(1/(\gamma^4 \epsilon^3))$ and achieve 0-1 error of $\eta+\epsilon$, where $\eta<1/2$ is the upper bound on the noise rate. Recent work gave evidence of an information-computation tradeoff, suggesting that a quadratic dependence on $1/\epsilon$ is required for computationally efficient algorithms. Our main result is a computationally efficient learner with sample complexity $\widetilde{\Theta}(1/(\gamma^2 \epsilon^2))$, nearly matching this lower bound. In addition, our algorithm is simple and practical, relying on online SGD on a carefully selected sequence of convex losses.
A machine learning tasks from observations must encounter and process uncertainty and novelty, especially when it is to maintain performance when observing new information and to choose the hypothesis that best fits the current observations. In this context, some key questions arise: what and how much information did the observations provide, how much information is required to identify the data-generating process, how many observations remain to get that information, and how does a predictor determine that it has observed novel information? This paper strengthens existing answers to these questions by formalizing the notion of identifiable information that arises from the language used to express the relationship between distinct states. Model identifiability and sample complexity are defined via computation of an indicator function over a set of hypotheses, bridging algorithmic and probabilistic information. Their properties and asymptotic statistics are described for data-generating processes ranging from deterministic processes to ergodic stationary stochastic processes. This connects the notion of identifying information in finite steps with asymptotic statistics and PAC-learning. The indicator function's computation naturally formalizes novel information and its identification from observations with respect to a hypothesis set. We also proved that computable PAC-Bayes learners' sample complexity distribution is determined by its moments in terms of the prior probability distribution over a fixed finite hypothesis set.
Machine learning algorithms often encounter different or "out-of-distribution" (OOD) data at deployment time, and OOD detection is frequently employed to detect these examples. While it works reasonably well in practice, existing theoretical results on OOD detection are highly pessimistic. In this work, we take a closer look at this problem, and make a distinction between uniform and non-uniform learnability, following PAC learning theory. We characterize under what conditions OOD detection is uniformly and non-uniformly learnable, and we show that in several cases, non-uniform learnability turns a number of negative results into positive. In all cases where OOD detection is learnable, we provide concrete learning algorithms and a sample-complexity analysis.
Monotone learning describes learning processes in which expected performance consistently improves as the amount of training data increases. However, recent studies challenge this conventional wisdom, revealing significant gaps in the understanding of generalization in machine learning. Addressing these gaps is crucial for advancing the theoretical foundations of the field. In this work, we utilize Probably Approximately Correct (PAC) learning theory to construct a theoretical risk distribution that approximates a learning algorithm's actual performance. We rigorously prove that this theoretical distribution exhibits monotonicity as sample sizes increase. We identify two scenarios under which deterministic algorithms based on Empirical Risk Minimization (ERM) are monotone: (1) the hypothesis space is finite, or (2) the hypothesis space has finite VC-dimension. Experiments on two classical learning problems validate our findings by demonstrating that the monotonicity of the algorithms' generalization error is guaranteed, as its theoretical risk upper bound monotonically converges to 0.
The learner's ability to generate a hypothesis that closely approximates the target function is crucial in machine learning. Achieving this requires sufficient data; however, unauthorized access by an eavesdropping learner can lead to security risks. Thus, it is important to ensure the performance of the "authorized" learner by limiting the quality of the training data accessible to eavesdroppers. Unlike previous studies focusing on encryption or access controls, we provide a theorem to ensure superior learning outcomes exclusively for the authorized learner with quantum label encoding. In this context, we use the probably-approximately-correct (PAC) learning framework and introduce the concept of learning probability to quantitatively assess learner performance. Our theorem allows the condition that, given a training dataset, an authorized learner is guaranteed to achieve a certain quality of learning outcome, while eavesdroppers are not. Notably, this condition can be constructed based only on the authorized-learning-only measurable quantities of the training data, i.e., its size and noise degree. We validate our theoretical proofs and predictions through convolutional neural networks (CNNs) image classification learning.
Cost-sensitive loss functions are crucial in many real-world prediction problems, where different types of errors are penalized differently; for example, in medical diagnosis, a false negative prediction can lead to worse consequences than a false positive prediction. However, traditional PAC learning theory has mostly focused on the symmetric 0-1 loss, leaving cost-sensitive losses largely unaddressed. In this work, we extend the celebrated theory of boosting to incorporate both cost-sensitive and multi-objective losses. Cost-sensitive losses assign costs to the entries of a confusion matrix, and are used to control the sum of prediction errors accounting for the cost of each error type. Multi-objective losses, on the other hand, simultaneously track multiple cost-sensitive losses, and are useful when the goal is to satisfy several criteria at once (e.g., minimizing false positives while keeping false negatives below a critical threshold). We develop a comprehensive theory of cost-sensitive and multi-objective boosting, providing a taxonomy of weak learning guarantees that distinguishes which guarantees are trivial (i.e., can always be achieved), which ones are boostable (i.e., imply strong learning), and which ones are intermediate, implying non-trivial yet not arbitrarily accurate learning. For binary classification, we establish a dichotomy: a weak learning guarantee is either trivial or boostable. In the multiclass setting, we describe a more intricate landscape of intermediate weak learning guarantees. Our characterization relies on a geometric interpretation of boosting, revealing a surprising equivalence between cost-sensitive and multi-objective losses.
We study the task of learning latent-variable models. A common algorithmic technique for this task is the method of moments. Unfortunately, moment-based approaches are hampered by the fact that the moment tensors of super-constant degree cannot even be written down in polynomial time. Motivated by such learning applications, we develop a general efficient algorithm for {\em implicit moment tensor computation}. Our framework generalizes the work of~\cite{LL21-opt} which developed an efficient algorithm for the specific moment tensors that arise in clustering mixtures of spherical Gaussians. By leveraging our implicit moment estimation algorithm, we obtain the first $\mathrm{poly}(d, k)$-time learning algorithms for the following models. * {\bf Mixtures of Linear Regressions} We give a $\mathrm{poly}(d, k, 1/\epsilon)$-time algorithm for this task, where $\epsilon$ is the desired error. * {\bf Mixtures of Spherical Gaussians} For density estimation, we give a $\mathrm{poly}(d, k, 1/\epsilon)$-time learning algorithm, where $\epsilon$ is the desired total variation error, under the condition that the means lie in a ball of radius $O(\sqrt{\log k})$. For parameter estimation, we give a $\mathrm{poly}(d, k, 1/\epsilon)$-time algorithm under the {\em optimal} mean separation of $\Omega(\log^{1/2}(k/\epsilon))$. * {\bf Positive Linear Combinations of Non-Linear Activations} We give a general algorithm for this task with complexity $\mathrm{poly}(d, k) g(\epsilon)$, where $\epsilon$ is the desired error and the function $g$ depends on the Hermite concentration of the target class of functions. Specifically, for positive linear combinations of ReLU activations, our algorithm has complexity $\mathrm{poly}(d, k) 2^{\mathrm{poly}(1/\epsilon)}$.
Delle Rose et al.~(COLT'23) introduced an effective version of the Vapnik-Chervonenkis dimension, and showed that it characterizes improper PAC learning with total computable learners. In this paper, we introduce and study a similar effectivization of the notion of Littlestone dimension. Finite effective Littlestone dimension is a necessary condition for computable online learning but is not a sufficient one -- which we already establish for classes of the effective Littlestone dimension 2. However, the effective Littlestone dimension equals the optimal mistake bound for computable learners in two special cases: a) for classes of Littlestone dimension 1 and b) when the learner receives as additional information an upper bound on the numbers to be guessed. Interestingly, finite effective Littlestone dimension also guarantees that the class consists only of computable functions.
Precision and Recall are foundational metrics in machine learning where both accurate predictions and comprehensive coverage are essential, such as in recommender systems and multi-label learning. In these tasks, balancing precision (the proportion of relevant items among those predicted) and recall (the proportion of relevant items successfully predicted) is crucial. A key challenge is that one-sided feedback--where only positive examples are observed during training--is inherent in many practical problems. For instance, in recommender systems like YouTube, training data only consists of videos that a user has actively selected, while unselected items remain unseen. Despite this lack of negative feedback in training, avoiding undesirable recommendations at test time is essential. We introduce a PAC learning framework where each hypothesis is represented by a graph, with edges indicating positive interactions, such as between users and items. This framework subsumes the classical binary and multi-class PAC learning models as well as multi-label learning with partial feedback, where only a single random correct label per example is observed, rather than all correct labels. Our work uncovers a rich statistical and algorithmic landscape, with nuanced boundaries on what can and cannot be learned. Notably, classical methods like Empirical Risk Minimization fail in this setting, even for simple hypothesis classes with only two hypotheses. To address these challenges, we develop novel algorithms that learn exclusively from positive data, effectively minimizing both precision and recall losses. Specifically, in the realizable setting, we design algorithms that achieve optimal sample complexity guarantees. In the agnostic case, we show that it is impossible to achieve additive error guarantees--as is standard in PAC learning--and instead obtain meaningful multiplicative approximations.
We revisit the problem of distribution learning within the framework of learning-augmented algorithms. In this setting, we explore the scenario where a probability distribution is provided as potentially inaccurate advice on the true, unknown distribution. Our objective is to develop learning algorithms whose sample complexity decreases as the quality of the advice improves, thereby surpassing standard learning lower bounds when the advice is sufficiently accurate. Specifically, we demonstrate that this outcome is achievable for the problem of learning a multivariate Gaussian distribution $N(\boldsymbol{\mu}, \boldsymbol{\Sigma})$ in the PAC learning setting. Classically, in the advice-free setting, $\tilde{\Theta}(d^2/\varepsilon^2)$ samples are sufficient and worst case necessary to learn $d$-dimensional Gaussians up to TV distance $\varepsilon$ with constant probability. When we are additionally given a parameter $\tilde{\boldsymbol{\Sigma}}$ as advice, we show that $\tilde{O}(d^{2-\beta}/\varepsilon^2)$ samples suffices whenever $\| \tilde{\boldsymbol{\Sigma}}^{-1/2} \boldsymbol{\Sigma} \tilde{\boldsymbol{\Sigma}}^{-1/2} - \boldsymbol{I_d} \|_1 \leq \varepsilon d^{1-\beta}$ (where $\|\cdot\|_1$ denotes the entrywise $\ell_1$ norm) for any $\beta > 0$, yielding a polynomial improvement over the advice-free setting.
We study the problem of PAC learning halfspaces in the reliable agnostic model of Kalai et al. (2012). The reliable PAC model captures learning scenarios where one type of error is costlier than the others. Our main positive result is a new algorithm for reliable learning of Gaussian halfspaces on $\mathbb{R}^d$ with sample and computational complexity $$d^{O(\log (\min\{1/\alpha, 1/\epsilon\}))}\min (2^{\log(1/\epsilon)^{O(\log (1/\alpha))}},2^{\mathrm{poly}(1/\epsilon)})\;,$$ where $\epsilon$ is the excess error and $\alpha$ is the bias of the optimal halfspace. We complement our upper bound with a Statistical Query lower bound suggesting that the $d^{\Omega(\log (1/\alpha))}$ dependence is best possible. Conceptually, our results imply a strong computational separation between reliable agnostic learning and standard agnostic learning of halfspaces in the Gaussian setting.
Accurate estimates of causal effects play a key role in decision-making across applications such as healthcare, economics, and operations. In the absence of randomized experiments, a common approach to estimating causal effects uses \textit{covariate adjustment}. In this paper, we study covariate adjustment for discrete distributions from the PAC learning perspective, assuming knowledge of a valid adjustment set $\bZ$, which might be high-dimensional. Our first main result PAC-bounds the estimation error of covariate adjustment by a term that is exponential in the size of the adjustment set; it is known that such a dependency is unavoidable even if one only aims to minimize the mean squared error. Motivated by this result, we introduce the notion of an \emph{$\eps$-Markov blanket}, give bounds on the misspecification error of using such a set for covariate adjustment, and provide an algorithm for $\eps$-Markov blanket discovery; our second main result upper bounds the sample complexity of this algorithm. Furthermore, we provide a misspecification error bound and a constraint-based algorithm that allow us to go beyond $\eps$-Markov blankets to even smaller adjustment sets. Our third main result upper bounds the sample complexity of this algorithm, and our final result combines the first three into an overall PAC bound. Altogether, our results highlight that one does not need to perfectly recover causal structure in order to ensure accurate estimates of causal effects.
In real-world applications, the distribution of the data, and our goals, evolve over time. The prevailing theoretical framework for studying machine learning, namely probably approximately correct (PAC) learning, largely ignores time. As a consequence, existing strategies to address the dynamic nature of data and goals exhibit poor real-world performance. This paper develops a theoretical framework called "Prospective Learning" that is tailored for situations when the optimal hypothesis changes over time. In PAC learning, empirical risk minimization (ERM) is known to be consistent. We develop a learner called Prospective ERM, which returns a sequence of predictors that make predictions on future data. We prove that the risk of prospective ERM converges to the Bayes risk under certain assumptions on the stochastic process generating the data. Prospective ERM, roughly speaking, incorporates time as an input in addition to the data. We show that standard ERM as done in PAC learning, without incorporating time, can result in failure to learn when distributions are dynamic. Numerical experiments illustrate that prospective ERM can learn synthetic and visual recognition problems constructed from MNIST and CIFAR-10. Code at https://github.com/neurodata/prolearn.
In this paper, we study the problem of methodically obtaining a sufficient set of kinesthetic demonstrations, one at a time, such that a robot can be confident of its ability to perform a complex manipulation task in a given region of its workspace. Although Learning from Demonstrations has been an active area of research, the problems of checking whether a set of demonstrations is sufficient, and systematically seeking additional demonstrations have remained open. We present a novel approach to address these open problems using (i) a screw geometric representation to generate manipulation plans from demonstrations, which makes the sufficiency of a set of demonstrations measurable; (ii) a sampling strategy based on PAC-learning from multi-armed bandit optimization to evaluate the robot's ability to generate manipulation plans in a subregion of its task space; and (iii) a heuristic to seek additional demonstration from areas of weakness. Thus, we present an approach for the robot to incrementally and actively ask for new demonstration examples until the robot can assess with high confidence that it can perform the task successfully. We present experimental results on two example manipulation tasks, namely, pouring and scooping, to illustrate our approach. A short video on the method: https://youtu.be/R-qICICdEos
This paper addresses the problem of PAC learning half spaces under constant malicious noise, where a fraction of the training data is adversarially corrupted. While traditional learning models assume clean data, real-world applications often face noisy environments that can significantly degrade the performance of machine learning algorithms. My study presents a novel, efficient algorithm that extends the existing theoretical frameworks to account for noise resilience in half space learning. By leveraging robust optimization techniques and advanced error-correction strategies, the proposed approach improves learning accuracy in adversarial conditions without requiring additional computational complexity. We provide a comprehensive analysis of the algorithm's performance, demonstrating its superior robustness to malicious noise when compared to existing state-of-the-art methods. Extensive theoretical evaluations are supported by empirical results that validate the algorithm's practical utility across a range of datasets and noise conditions. This work contributes to the field by offering a new, scalable solution to learning under noise, enhancing the reliability of machine learning systems in adversarial settings.