Search for: All records

We study the problem of agnostic PAC reinforcement learning (RL): given a policy class Pi, how many rounds of interaction with an unknown MDP (with a potentially large state and action space) are required to learn an epsilon-suboptimal policy with respect to Pi? Towards that end, we introduce a new complexity measure, called the spanning capacity, that depends solely on the set Pi and is independent of the MDP dynamics. With a generative model, we show that the spanning capacity characterizes PAC learnability for every policy class Pi. However, for online RL, the situation is more subtle. We show there exists a policy class Pi with a bounded spanning capacity that requires a superpolynomial number of samples to learn. This reveals a surprising separation for agnostic learnability between generative access and online access models (as well as between deterministic/stochastic MDPs under online access). On the positive side, we identify an additional sunflower structure which in conjunction with bounded spanning capacity enables statistically efficient online RL via a new algorithm called POPLER, which takes inspiration from classical importance sampling methods as well as recent developments for reachable-state identification and policy evaluation in reward-free exploration. 

The susceptibility of modern machine learning classifiers to adversarial examples has motivated theoretical results suggesting that these might be unavoidable. However, these results can be too general to be applicable to natural data distributions. Indeed, humans are quite robust for tasks involving vision. This apparent conflict motivates a deeper dive into the question: Are adversarial examples truly unavoidable? In this work, we theoretically demonstrate that a key property of the data distribution – concentration on small-volume subsets of the input space – determines whether a robust classifier exists. We further demonstrate that, for a data distribution concentrated on a union of low-dimensional linear subspaces, utilizing structure in data naturally leads to classifiers that enjoy data-dependent polyhedral robustness guarantees, improving upon methods for provable certification in certain regimes. 

Graph neural networks (GNNs) are commonly described as being permutation equivariant with respect to node relabeling in the graph. This symmetry of GNNs is often compared to the translation equivariance of Euclidean convolution neural networks (CNNs). However, these two symmetries are fundamentally different: The translation equivariance of CNNs corresponds to symmetries of the fixed domain acting on the image signals (sometimes known as active symmetries), whereas in GNNs any permutation acts on both the graph signals and the graph domain (sometimes described as passive symmetries). In this work, we focus on the active symmetries of GNNs, by considering a learning setting where signals are supported on a fixed graph. In this case, the natural symmetries of GNNs are the automorphisms of the graph. Since real-world graphs tend to be asymmetric, we relax the notion of symmetries by formalizing approximate symmetries via graph coarsening. We present a bias-variance formula that quantifies the tradeoff between the loss in expressivity and the gain in the regularity of the learned estimator, depending on the chosen symmetry group. To illustrate our approach, we conduct extensive experiments on image inpainting, traffic flow prediction, and human pose estimation with different choices of symmetries. We show theoretically and empirically that the best generalization performance can be achieved by choosing a suitably larger group than the graph automorphism, but smaller than the permutation group. 

Multi-agent reinforcement learning (MARL) has primarily focused on solving a single task in isolation, while in practice the environment is often evolving, leaving many related tasks to be solved. In this paper, we investigate the benefits of meta-learning in solving multiple MARL tasks collectively. We establish the first line of theoretical results for meta-learning in a wide range of fundamental MARL settings, including learning Nash equilibria in two-player zero-sum Markov games and Markov potential games, as well as learning coarse correlated equilibria in general-sum Markov games. Under natural notions of task similarity, we show that meta-learning achieves provable sharper convergence to various game-theoretical solution concepts than learning each task separately. As an important intermediate step, we develop multiple MARL algorithms with initialization-dependent convergence guarantees. Such algorithms integrate optimistic policy mirror descents with stage-based value updates, and their refined convergence guarantees (nearly) recover the best known results even when a good initialization is unknown. To our best knowledge, such results are also new and might be of independent interest. We further provide numerical simulations to corroborate our theoretical findings. 

Learning from noisy labels is a long-standing problem in machine learning for real applications. One of the main research lines focuses on learning a label corrector to purify potential noisy labels. However, these methods typically rely on strict assumptions and are limited to certain types of label noise. In this paper, we reformulate the label-noise problem from a generative-model perspective, i.e., labels are generated by gradually refining an initial random guess. This new perspective immediately enables existing powerful diffusion models to seamlessly learn the stochastic generative process. Once the generative uncertainty is modeled, we can perform classification inference using maximum likelihood estimation of labels. To mitigate the impact of noisy labels, we propose Label-Retrieval- Augmented (LRA) diffusion model 1, which leverages neighbor consistency to effectively construct pseudo-clean labels for diffusion training. Our model is flexible and general, allowing easy incorporation of different types of conditional information, e.g., use of pre-trained models, to further boost model performance. Extensive experiments are conducted for evaluation. Our model achieves new state-of-the-art (SOTA) results on all standard real-world benchmark datasets. Remarkably, by incorporating conditional information from the powerful CLIP model, our method can boost the current SOTA accuracy by 10-20 absolute points in many cases. 

We consider the problem of learning a single-index target function f∗ : Rd → R under the spiked covariance data: f∗(x) = σ∗   √ 1 1+θ ⟨x,μ⟩   , x ∼ N(0, Id + θμμ⊤), θ ≍ dβ for β ∈ [0, 1), where the link function σ∗ : R → R is a degree-p polynomial with information exponent k (defined as the lowest degree in the Hermite expansion of σ∗), and it depends on the projection of input x onto the spike (signal) direction μ ∈ Rd. In the proportional asymptotic limit where the number of training examples n and the dimensionality d jointly diverge: n, d → ∞, n/d → ψ ∈ (0,∞), we ask the following question: how large should the spike magnitude θ be, in order for (i) kernel methods, (ii) neural networks optimized by gradient descent, to learn f∗? We show that for kernel ridge regression, β ≥ 1 − 1 p is both sufficient and necessary. Whereas for two-layer neural networks trained with gradient descent, β > 1 − 1 k suffices. Our results demonstrate that both kernel methods and neural networks benefit from low-dimensional structures in the data. Further, since k ≤ p by definition, neural networks can adapt to such structures more effectively. 

Given a set of points of interest, a volumetric spanner is a subset of the points using which all the points can be expressed using “small” coefficients (measured in an appropriate norm). This notion, which has also been referred to as a well-conditioned basis, has found several applications, including bandit linear optimization, determinant maximization, and matrix low rank approximation. In this paper, we give almost optimal bounds on the size of volumetric spanners for all L_p norms, and show that they can be constructed using a simple local search procedure. We then show the applications of our result to other tasks and in particular the problem of finding coresets for the Minimum Volume Enclosing Ellipsoid (MVEE) problem. 

This paper presents a unified approach for maximizing continuous DR-submodular functions that encompasses a range of settings and oracle access types. Our approach includes a Frank-Wolfe type offline algorithm for both monotone and non-monotone functions, with different restrictions on the general convex set. We consider settings where the oracle provides access to either the gradient of the function or only the function value, and where the oracle access is either deterministic or stochastic. We determine the number of required oracle accesses in all cases. Our approach gives new/improved results for nine out of the sixteen considered cases, avoids computationally expensive projections in three cases, with the proposed framework matching performance of state-of-the-art approaches in the remaining four cases. Notably, our approach for the stochastic function value-based oracle enables the first regret bounds with bandit feedback for stochastic DR-submodular functions. 

We investigate replicable learning algorithms. Informally a learning algorithm is replicable if the algorithm outputs the same canonical hypothesis over multiple runs with high probability, even when different runs observe a different set of samples from the unknown data distribution. In general, such a strong notion of replicability is not achievable. Thus we consider two feasible notions of replicability called {\em list replicability} and {\em certificate replicability}. Intuitively, these notions capture the degree of (non) replicability. The goal is to design learning algorithms with optimal list and certificate complexities while minimizing the sample complexity. Our contributions are the following. 1. We first study the learning task of estimating the biases of $$d$$ coins, up to an additive error of $$\varepsilon$$, by observing samples. For this task, we design a $(d+1)$-list replicable algorithm. To complement this result, we establish that the list complexity is optimal, i.e there are no learning algorithms with a list size smaller than $d+1$ for this task. We also design learning algorithms with certificate complexity $$\tilde{O}(\log d)$$. The sample complexity of both these algorithms is $$\tilde{O}(\frac{d^2}{\varepsilon^2})$$ where $$\varepsilon$$ is the approximation error parameter (for a constant error probability). 2. In the PAC model, we show that any hypothesis class that is learnable with $$d$$-nonadaptive statistical queries can be learned via a $(d+1)$-list replicable algorithm and also via a $$\tilde{O}(\log d)$$-certificate replicable algorithm. The sample complexity of both these algorithms is $$\tilde{O}(\frac{d^2}{\nu^2})$$ where $$\nu$$ is the approximation error of the statistical query. We also show that for the concept class \dtep, the list complexity is exactly $d+1$ with respect to the uniform distribution. To establish our upper bound results we use rounding schemes induced by geometric partitions with certain properties. We use Sperner/KKM Lemma to establish the lower bound results. 

We consider distributed parameter estimation using interactive protocols subject to local information constraints such as bandwidth limitations, local differential privacy, and restricted measurements. We provide a unified framework enabling us to derive a variety of (tight) minimax lower bounds for different parametric families of distributions, both continuous and discrete, under any Lp loss. Our lower bound framework is versatile and yields “plug-and-play” bounds that are widely applicable to a large range of estimation problems, and, for the prototypical case of the Gaussian family, circumvents limitations of previous techniques. In particular, our approach recovers bounds obtained using data processing inequalities and Cramér–Rao bounds, two other alternative approaches for proving lower bounds in our setting of interest. Further, for the families considered, we complement our lower bounds with matching upper bounds.