Most value function learning algorithms in reinforcement learning are based on the mean squared (projected) Bellman error. However, squared errors are known to be sensitive to outliers, both skewing the solution of the objective and resulting in high-magnitude and high-variance gradients. To control these high-magnitude updates, typical strategies in RL involve clipping gradients, clipping rewards, rescaling rewards, or clipping errors. While these strategies appear to be related to robust losses -- like the Huber loss -- they are built on semi-gradient update rules which do not minimize a known loss. In this work, we build on recent insights reformulating squared Bellman errors as a saddlepoint optimization problem and propose a saddlepoint reformulation for a Huber Bellman error and Absolute Bellman error. We start from a formalization of robust losses, then derive sound gradient-based approaches to minimize these losses in both the online off-policy prediction and control settings. We characterize the solutions of the robust losses, providing insight into the problem settings where the robust losses define notably better solutions than the mean squared Bellman error. Finally, we show that the resulting gradient-based algorithms are more stable, for both prediction and control, with less sensitivity to meta-parameters.

In this paper, we investigate a distributed aggregative optimization problem in a network, where each agent has its own local cost function which depends not only on the local state variable but also on an aggregated function of state variables from all agents. To accelerate the optimization process, we combine heavy ball and Nesterov's accelerated methods with distributed aggregative gradient tracking, and propose two novel algorithms named DAGT-HB and DAGT-NES for solving the distributed aggregative optimization problem. We analyse that the DAGT-HB and DAGT-NES algorithms can converge to an optimal solution at a global $\mathbf{R}-$linear convergence rate when the objective function is smooth and strongly convex, and when the parameters (e.g., step size and momentum coefficients) are selected within certain ranges. A numerical experiment on the optimal placement problem is given to verify the effectiveness and superiority of our proposed algorithms.

Despite their great success in image recognition tasks, deep neural networks (DNNs) have been observed to be susceptible to universal adversarial perturbations (UAPs) which perturb all input samples with a single perturbation vector. However, UAPs often struggle in transferring across DNN architectures and lead to challenging optimization problems. In this work, we study the transferability of UAPs by analyzing equilibrium in the universal adversarial example game between the classifier and UAP adversary players. We show that under mild assumptions the universal adversarial example game lacks a pure Nash equilibrium, indicating UAPs' suboptimal transferability across DNN classifiers. To address this issue, we propose Universal Adversarial Directions (UADs) which only fix a universal direction for adversarial perturbations and allow the perturbations' magnitude to be chosen freely across samples. We prove that the UAD adversarial example game can possess a Nash equilibrium with a pure UAD strategy, implying the potential transferability of UADs. We also connect the UAD optimization problem to the well-known principal component analysis (PCA) and develop an efficient PCA-based algorithm for optimizing UADs. We evaluate UADs over multiple benchmark image datasets. Our numerical results show the superior transferability of UADs over standard gradient-based UAPs.

ABSTRACT Speech-centric machine learning systems have revolutionized a number of leading industries ranging from transportation and healthcare to education and defense, fundamentally reshaping how people live, work, and interact with each other. However, recent studies have demonstrated that many speech-centric ML systems may need to be considered more trustworthy for broader deployment. Specifically, concerns over privacy breaches, discriminating performance, and vulnerability to adversarial attacks have all been discovered in ML research fields. In order to address the above challenges and risks, a significant number of efforts have been made to ensure these ML systems are trustworthy, especially private, safe, and fair. In this paper, we conduct the first comprehensive survey on speech-centric trustworthy ML topics related to privacy, safety, and fairness. In addition to serving as a summary report for the research community, we highlight several promising future research directions to inspire researchers who wish to explore further in this area.

This paper proposes Differential-Critic Generative Adversarial Network (DiCGAN) to learn the distribution of user-desired data when only partial instead of the entire dataset possesses the desired property. DiCGAN generates desired data that meets the user's expectations and can assist in designing biological products with desired properties. Existing approaches select the desired samples first and train regular GANs on the selected samples to derive the user-desired data distribution. However, the selection of the desired data relies on global knowledge and supervision over the entire dataset. DiCGAN introduces a differential critic that learns from pairwise preferences, which are local knowledge and can be defined on a part of training data. The critic is built by defining an additional ranking loss over the Wasserstein GAN's critic. It endows the difference of critic values between each pair of samples with the user preference and guides the generation of the desired data instead of the whole data. For a more efficient solution to ensure data quality, we further reformulate DiCGAN as a constrained optimization problem, based on which we theoretically prove the convergence of our DiCGAN. Extensive experiments on a diverse set of datasets with various applications demonstrate that our DiCGAN achieves state-of-the-art performance in learning the user-desired data distributions, especially in the cases of insufficient desired data and limited supervision.

Contemporary applications of machine learning in two-team e-sports and the superior expressivity of multi-agent generative adversarial networks raise important and overlooked theoretical questions regarding optimization in two-team games. Formally, two-team zero-sum games are defined as multi-player games where players are split into two competing sets of agents, each experiencing a utility identical to that of their teammates and opposite to that of the opposing team. We focus on the solution concept of Nash equilibria (NE). We first show that computing NE for this class of games is $\textit{hard}$ for the complexity class ${\mathrm{CLS}}$. To further examine the capabilities of online learning algorithms in games with full-information feedback, we propose a benchmark of a simple -- yet nontrivial -- family of such games. These games do not enjoy the properties used to prove convergence for relevant algorithms. In particular, we use a dynamical systems perspective to demonstrate that gradient descent-ascent, its optimistic variant, optimistic multiplicative weights update, and extra gradient fail to converge (even locally) to a Nash equilibrium. On a brighter note, we propose a first-order method that leverages control theory techniques and under some conditions enjoys last-iterate local convergence to a Nash equilibrium. We also believe our proposed method is of independent interest for general min-max optimization.

Federated learning (FL) aims to minimize the communication complexity of training a model over heterogeneous data distributed across many clients. A common approach is local methods, where clients take multiple optimization steps over local data before communicating with the server (e.g., FedAvg). Local methods can exploit similarity between clients' data. However, in existing analyses, this comes at the cost of slow convergence in terms of the dependence on the number of communication rounds R. On the other hand, global methods, where clients simply return a gradient vector in each round (e.g., SGD), converge faster in terms of R but fail to exploit the similarity between clients even when clients are homogeneous. We propose FedChain, an algorithmic framework that combines the strengths of local methods and global methods to achieve fast convergence in terms of R while leveraging the similarity between clients. Using FedChain, we instantiate algorithms that improve upon previously known rates in the general convex and PL settings, and are near-optimal (via an algorithm-independent lower bound that we show) for problems that satisfy strong convexity. Empirical results support this theoretical gain over existing methods.

Intelligent Virtual Machine (VM) provisioning is central to cost and resource efficient computation in cloud computing environments. As bootstrapping VMs is time-consuming, a key challenge for latency-critical tasks is to predict future workload demands to provision VMs proactively. However, existing AI-based solutions tend to not holistically consider all crucial aspects such as provisioning overheads, heterogeneous VM costs and Quality of Service (QoS) of the cloud system. To address this, we propose a novel method, called CILP, that formulates the VM provisioning problem as two sub-problems of prediction and optimization, where the provisioning plan is optimized based on predicted workload demands. CILP leverages a neural network as a surrogate model to predict future workload demands with a co-simulated digital-twin of the infrastructure to compute QoS scores. We extend the neural network to also act as an imitation learner that dynamically decides the optimal VM provisioning plan. A transformer based neural model reduces training and inference overheads while our novel two-phase decision making loop facilitates in making informed provisioning decisions. Crucially, we address limitations of prior work by including resource utilization, deployment costs and provisioning overheads to inform the provisioning decisions in our imitation learning framework. Experiments with three public benchmarks demonstrate that CILP gives up to 22% higher resource utilization, 14% higher QoS scores and 44% lower execution costs compared to the current online and offline optimization based state-of-the-art methods.

Structural learning, which aims to learn directed acyclic graphs (DAGs) from observational data, is foundational to causal reasoning and scientific discovery. Recent advancements formulate structural learning into a continuous optimization problem; however, DAG learning remains a highly non-convex problem, and there has not been much work on leveraging well-developed convex optimization techniques for causal structural learning. We fill this gap by proposing a data-adaptive linear approach for causal structural learning from time series data, which can be conveniently cast into a convex optimization problem using a recently developed monotone operator variational inequality (VI) formulation. Furthermore, we establish non-asymptotic recovery guarantee of the VI-based approach and show the superior performance of our proposed method on structure recovery over existing methods via extensive numerical experiments.

The accuracy and complexity of machine learning algorithms based on kernel optimization are determined by the set of kernels over which they are able to optimize. An ideal set of kernels should: admit a linear parameterization (for tractability); be dense in the set of all kernels (for robustness); be universal (for accuracy). Recently, a framework was proposed for using positive matrices to parameterize a class of positive semi-separable kernels. Although this class can be shown to meet all three criteria, previous algorithms for optimization of such kernels were limited to classification and furthermore relied on computationally complex Semidefinite Programming (SDP) algorithms. In this paper, we pose the problem of learning semiseparable kernels as a minimax optimization problem and propose a SVD-QCQP primal-dual algorithm which dramatically reduces the computational complexity as compared with previous SDP-based approaches. Furthermore, we provide an efficient implementation of this algorithm for both classification and regression -- an implementation which enables us to solve problems with 100 features and up to 30,000 datums. Finally, when applied to benchmark data, the algorithm demonstrates the potential for significant improvement in accuracy over typical (but non-convex) approaches such as Neural Nets and Random Forest with similar or better computation time.