AISTATS 2023 Papers

Decentralized learning offers privacy and communication efficiency when data are naturally distributed among agents communicating over an underlying graph. Motivated by overparameterized learning settings, in which models are trained to zero training loss, we study algorithmic and generalization properties of decentralized learning with gradient descent on separable data. Specifically, for decentralized gradient descent (DGD) and a variety of loss functions that asymptote to zero at infinity (including exponential and logistic losses), we derive novel finite-time generalization bounds. This complements a long line of recent work that studies the generalization performance and the implicit bias of gradient descent over separable data, but has thus far been limited to centralized learning scenarios. Notably, our generalization bounds approximately match in order their centralized counterparts. Critical behind this, and of independent interest, is establishing novel bounds on the training loss and the rate-of-consensus of DGD for a class of self-bounded losses. Finally, on the algorithmic front, we design improved gradient-based routines for decentralized learning with separable data and empirically demonstrate orders-of-magnitude of speed-up in terms of both training and generalization performance.

We study the problem of reconstructing a continuous multidimensional signal from a small number of samples under Fourier constraints assuming that the Fourier power spectrum of the signal has some desirable properties, e.g. being compactly supported, being sparse. We further assume that the Fourier constraint can be expressed as a prior distribution on the Fourier power spectrum, which subsumes the aforementioned examples. The study of sampling and reconstructing in this vein has attracted much attention with a long history. In this paper, we are interested in finding oblivious sampling strategies, that is, sampling without knowing what specific constraint is put on the Fourier power spectrum. We show that it is possible to obliviously sample a Fourier-constrained multidimensional signal with a near-optimal (up to a logarithmic factor) number of samples that guarantee successful reconstruction, partially answering an open question in Avron et al. (2019) which considered the $1$-dimensional case. Our approach highlights a phenomenon that is unique for dimension $d\ge 2$ that the sampling strategy should depend on the geometry of the region on which the signal is to be reconstructed, unlike the case $d=1$ where all regions are of the form $[a,b]$ which are all geometrically equivalent. Our proof, using tools from convex geometry, also illuminates an idea obscured in $d=1$, that to reconstruct a signal in a given region, it can be helpful to take some samples outside that region.

The essential difficulty of gradient-based bilevel optimization using implicit differentiation is to estimate the inverse Hessian vector product with respect to neural network parameters. This paper proposes to tackle this problem by the Nyström method and the Woodbury matrix identity, exploiting the low-rankness of the Hessian. Compared to existing methods using iterative approximation, such as conjugate gradient and the Neumann series approximation, the proposed method avoids numerical instability and can be efficiently computed in matrix operations without iterations. As a result, the proposed method works stably in various tasks and is faster than iterative approximations. Throughout experiments including large-scale hyperparameter optimization and meta learning, we demonstrate that the Nyström method consistently achieves comparable or even superior performance to other approaches. The source code is available from https://github.com/moskomule/hypergrad.

In critical applications, causal models are the prime choice for their trustworthiness and explainability. If data is inherently distributed and privacy-sensitive, federated learning allows for collaboratively training a joint model. Existing approaches for federated causal discovery share locally discovered causal model in every iteration, therewith not only revealing local structure but also leading to very high communication costs. Instead, we propose an approach for privacy-preserving federated causal discovery by distributed min-max regret optimization. We prove that max-regret is a consistent scoring criterion that can be used within the well-known Greedy Equivalence Search to discover causal networks in a federated setting and is provably privacy-preserving at the same time. Through extensive experiments, we show that our approach reliably discovers causal networks without ever looking at local data and beats the state of the art both in terms of the quality of discovered causal networks as well as communication efficiency.

We study a contextual version of online combinatorial optimisation with full and semi-bandit feedback. In this sequential decision-making problem, an online learner has to select an action from a combinatorial decision space after seeing a vector-valued context in each round. As a result of its action, the learner incurs a loss that is a bilinear function of the context vector and the vector representation of the chosen action. We consider two natural versions of the problem: semi-bandit where the losses are revealed for each component appearing in the learner’s combinatorial action, and full-bandit where only the total loss is observed. We design computationally efficient algorithms based on a new loss estimator that takes advantage of the special structure of the problem, and show regret bounds order $\sqrt{T}$ with respect to the time horizon. The bounds demonstrate polynomial scaling with the relevant problem parameters which is shown to be nearly optimal. The theoretical results are complemented by a set of experiments on simulated data.

Thompson sampling has proven effective across a wide range of stationary bandit environments. However, as we demonstrate in this paper, it can perform poorly when applied to nonstationary environments. We show that such failures are attributed to the fact that, when exploring, the algorithm does not differentiate actions based on how quickly the information acquired loses its usefulness due to nonstationarity. Building upon this insight, we propose predictive sampling, an algorithm that deprioritizes acquiring information that quickly loses usefulness. Theoretical guarantee on the performance of predictive sampling is established through a Bayesian regret bound. We provide versions of predictive sampling for which computations tractably scale to complex bandit environments of practical interest. Through numerical simulation, we demonstrate that predictive sampling outperforms Thompson sampling in all nonstationary environments examined.

Typical models of active learning assume a learner can directly manipulate or query a covariate X to study its relationship with a response Y. However, if X is a feature of a complex system, it may be possible only to indirectly influence X by manipulating a control variable Z, a scenario we refer to as Indirect Active Learning. Under a nonparametric fixed-budget model of Indirect Active Learning, we study minimax convergence rates for estimating a local relationship between X and Y, with different rates depending on the complexities and noise levels of the relationships between Z and X and between X and Y. We also derive minimax rates for passive learning under comparable assumptions, finding in many cases that, while there is an asymptotic benefit to active learning, this benefit is fully realized by a simple two-stage learner that runs two passive experiments in sequence.

A key challenge in the practical application of Gaussian processes (GPs) is selecting a proper covariance function. The process convolutions construction of GPs allows some additional flexibility, but still requires choosing a proper smoothing kernel, which is non-trivial. Previous approaches have built covariance functions by using GP priors over the smoothing kernel, and by extension the covariance, as a way to bypass the need to specify it in advance. However, these models have been limited in several ways: they are restricted to single dimensional inputs, e.g. time; they only allow modelling of single outputs and they do not scale to large datasets since inference is not straightforward. In this paper, we introduce a nonparametric process convolution formulation for GPs that alleviates these weaknesses. We achieve this using a functional sampling approach based on Matheron’s rule to perform fast sampling using interdomain inducing variables. We test the performance of our model on benchmarks for single output, multi-output and large-scale GP regression, and find that our approach can provide improvements over standard GP models, particularly for larger datasets.

Active search is a setting in adaptive experimental design where we aim to uncover members of rare, valuable class(es) subject to a budget constraint. An important consideration in this problem is diversity among the discovered targets – in many applications, diverse discoveries offer more insight and may be preferable in downstream tasks. However, most existing active search policies either assume that all targets belong to a common positive class or encourage diversity via simple heuristics. We present a novel formulation of active search with multiple target classes, characterized by a utility function chosen from a flexible family whose members encourage diversity among discoveries via a diminishing returns mechanism. We then study this problem under the Bayesian lens and prove a hardness result for approximating the optimal policy for arbitrary positive, increasing, and concave utility functions. Finally, we design an efficient, nonmyopic approximation to the optimal policy for this class of utilities and demonstrate its superior empirical performance in a variety of experimental settings, including drug discovery.

This paper is concerned with low-rank matrix optimization, which has found a wide range of applications in machine learning. This problem in the special case of matrix sensing has been studied extensively through the notion of Restricted Isometry Property (RIP), leading to a wealth of results on the geometric landscape of the problem and the convergence rate of common algorithms. However, the existing results can handle the problem in the case with a general objective function subject to noisy data only when the RIP constant is close to 0. In this paper, we develop a new mathematical framework to solve the above-mentioned problem with a far less restrictive RIP constant. We prove that as long as the RIP constant of the noiseless objective is less than 1/3, any spurious local solution of the noisy optimization problem must be close to the ground truth solution. By working through the strict saddle property, we also show that an approximate solution can be found in polynomial time. We characterize the geometry of the spurious local minima of the problem in a local region around the ground truth in the case when the RIP constant is greater than 1/3. Compared to the existing results in the literature, this paper offers the strongest RIP bound, and provides a complete theoretical analysis on the global and local optimization landscapes of general low-rank optimization problems under random corruptions from any finite-variance family.

While generation of synthetic data under differential privacy (DP) has received a lot of attention in the data privacy community, analysis of synthetic data has received much less. Existing work has shown that simply analysing DP synthetic data as if it were real does not produce valid inferences of population-level quantities. For example, confidence intervals become too narrow, which we demonstrate with a simple experiment. We tackle this problem by combining synthetic data analysis techniques from the field of multiple imputation (MI), and synthetic data generation using noise-aware (NA) Bayesian modeling into a pipeline NA+MI that allows computing accurate uncertainty estimates for population-level quantities from DP synthetic data. To implement NA+MI for discrete data generation using the values of marginal queries, we develop a novel noise-aware synthetic data generation algorithm NAPSU-MQ using the principle of maximum entropy. Our experiments demonstrate that the pipeline is able to produce accurate confidence intervals from DP synthetic data. The intervals become wider with tighter privacy to accurately capture the additional uncertainty stemming from DP noise.

The Nash equilibrium (NE) is a classic solution concept for normal-form games that is stable under potential unilateral deviations by self-interested agents. Bayesian optimization (BO) has been used to find NE in continuous general-sum games with unknown costly-to-sample utility functions in a sample-efficient manner. This paper presents the first no-regret BO algorithm that is sample-efficient in finding pure NE by leveraging theory on high probability confidence bounds with Gaussian processes and the maximum information gain of kernel functions. Unlike previous works, our algorithm is theoretically guaranteed to converge to the optimal solution (i.e., NE). We also introduce the novel setting of applying BO to finding mixed NE in unknown discrete general-sum games and show that our theoretical framework is general enough to be extended naturally to this setting by developing a no-regret BO algorithm that is sample-efficient in finding mixed NE. We empirically show that our algorithms are competitive w.r.t. suitable baselines in finding NE.

Supply chain management (SCM) has been recognized as an important discipline with applications to many industries, where the two-echelon stochastic inventory model, involving one downstream retailer and one upstream supplier, plays a fundamental role for developing firms’ SCM strategies. In this work, we aim at designing online learning algorithms for this problem with an unknown demand distribution, which brings distinct features as compared to classic online convex optimization problems. Specifically, we consider the two-echelon supply chain model introduced in [Cachon and Zipkin, 1999] under two different settings: the centralized setting, where a planner decides both agents’ strategy simultaneously, and the decentralized setting, where two agents decide their strategy independently and selfishly. We design algorithms that achieve favorable guarantees for both regret and convergence to the optimal inventory decision in both settings, and additionally for individual regret in the decentralized setting. Our algorithms are based on Online Gradient Descent and Online Newton Step, together with several new ingredients specifically designed for our problem. We also implement our algorithms and show their empirical effectiveness.

Pandemics have a major impact on society and the economy. In the case of a new virus, such as COVID-19, high-grade tests and vaccines might be slow to develop and scarce in the crucial initial phase. With no time to waste and lock-downs being expensive, contact tracing is thus an essential tool for policymakers. In theory, statistical inference on a virus transmission model can provide an effective method for tracing infections. However, in practice, such algorithms need to run decentralized, rendering existing methods – that require hundreds or even thousands of daily messages per person – infeasible. In this paper, we develop an algorithm that (i) requires only a few (2-5) daily messages, (ii) works with extremely low bandwidths (3-5 bits) and (iii) enables quarantining and targeted testing that drastically reduces the peak and length of the pandemic. We compare the effectiveness of our algorithm using two agent-based simulators of realistic contact patterns and pandemic parameters and show that it performs well even with low bandwidth, imprecise tests, and incomplete population coverage.

Simulated annealing (SA) is a stochastic global optimisation metaheuristic applicable to a wide range of discrete and continuous variable problems. Despite its simplicity, SA hinges on carefully handpicked components, viz. proposal distribution and annealing schedule, that often have to be fine tuned to individual problem instances. In this work, we seek to make SA more effective and easier to use by framing its proposal distribution as a reinforcement learning policy that can be optimised for higher solution quality given a computational budget. The result is Neural SA, a competitive and general machine learning method for combinatorial optimisation that is efficient, and easy to design and train. We show Neural SA with such a learnt proposal distribution, parametrised by small equivariant neural networks, outperforms SA baselines on several problems: Rosenbrock’s function and the Knapsack, Bin Packing and Travelling Salesperson problems. We also show Neural SA scales well to large problems (generalising to much larger instances than those seen during training) while getting comparable performance to popular off-the-shelf solvers and machine learning methods in terms of solution quality and wall-clock time.

Many real-world offline reinforcement learning (RL) problems involve continuous-time environments with delays. Such environments are characterized by two distinctive features: firstly, the state x(t) is observed at irregular time intervals, and secondly, the current action a(t) only affects the future state x(t + g) with an unknown delay g > 0. A prime example of such an environment is satellite control where the communication link between earth and a satellite causes irregular observations and delays. Existing offline RL algorithms have achieved success in environments with irregularly observed states in time or known delays. However, environments involving both irregular observations in time and unknown delays remains an open and challenging problem. To this end, we propose Neural Laplace Control, a continuous-time model-based offline RL method that combines a Neural Laplace dynamics model with a model predictive control (MPC) planner–and is able to learn from an offline dataset sampled with irregular time intervals from an environment that has a inherent unknown constant delay. We show experimentally on continuous-time delayed environments it is able to achieve near expert policy performance.

We consider the problem of discovering subgroup $H$ of permutation group $S_n$. Unlike the traditional $H$-invariant networks wherein $H$ is assumed to be known, we present a method to discover the underlying subgroup, given that it satisfies certain conditions. Our results show that one could discover any subgroup of type $S_k (k \leq n)$ by learning an $S_n$-invariant function and a linear transformation. We also prove similar results for cyclic and dihedral subgroups. Finally, we provide a general theorem that can be extended to discover other subgroups of $S_n$. We also demonstrate the applicability of our results through numerical experiments on image-digit sum and symmetric polynomial regression tasks.

We consider the problem of model estimation in episodic Block MDPs. In these MDPs, the decision maker has access to rich observations or contexts generated from a small number of latent states. We are interested in estimating the latent state decoding function (the mapping from the observations to latent states) based on data generated under a fixed behavior policy. We derive an information-theoretical lower bound on the error rate for estimating this function and present an algorithm approaching this fundamental limit. In turn, our algorithm also provides estimates of all the components of the MDP. We apply our results to the problem of learning near-optimal policies in the reward-free setting. Based on our efficient model estimation algorithm, we show that we can infer a policy converging (as the number of collected samples grows large) to the optimal policy at the best possible rate. Our analysis provides necessary and sufficient conditions under which exploiting the block structure yields improvements in the sample complexity for identifying near-optimal policies. When these conditions are met, the sample complexity in the minimax reward-free setting is improved by a multiplicative factor $n$, where $n$ is the number of possible contexts.

Motivated by personalized healthcare and other applications involving sensitive data, we study online exploration in reinforcement learning with differential privacy (DP) constraints. Existing work on this problem established that no-regret learning is possible under joint differential privacy (JDP) and local differential privacy (LDP) but did not provide an algorithm with optimal regret. We close this gap for the JDP case by designing an $\epsilon$-JDP algorithm with a regret of $\widetilde{O}(\sqrt{SAH^2T}+S^2AH^3/\epsilon)$ which matches the information-theoretic lower bound of non-private learning for all choices of $\epsilon> S^{1.5}A^{0.5} H^2/\sqrt{T}$. In the above, $S$, $A$ denote the number of states and actions, $H$ denotes the planning horizon, and $T$ is the number of steps. To the best of our knowledge, this is the first private RL algorithm that achieves privacy for free asymptotically as $T\rightarrow \infty$. Our techniques — which could be of independent interest — include privately releasing Bernstein-type exploration bonuses and an improved method for releasing visitation statistics. The same techniques also imply a slightly improved regret bound for the LDP case.