In this paper, we study the minimax optimization problem in the smooth and strongly convex-strongly concave setting when we have access to noisy estimates of gradients. In particular, we first analyze the stochastic Gradient Descent Ascent (GDA) method with constant stepsize, and show that it converges to a neighborhood of the solution of the minimax problem. We further provide tight bounds on the convergence rate and the size of this neighborhood. Next, we propose a multistage variant of stochastic GDA (M-GDA) that runs in multiple stages with a particular learning rate decay schedule and converges to the exact solution of the minimax problem. We show M-GDA achieves the lower bounds in terms of noise dependence without any assumptions on the knowledge of noise characteristics. We also show that M-GDA obtains a linear decay rate with respect to the error's dependence on the initial error, although the dependence on condition number is suboptimal. In order to improve this dependence, we apply the multistage machinery to the stochastic Optimistic Gradient Descent Ascent (OGDA) algorithm and propose the M-OGDA algorithm which also achieves the optimal linear decay rate with respect to the initial error. To the best of our knowledge, this method is the first to simultaneously achieve the best dependence on noise characteristic as well as the initial error and condition number.

Much work on social media opinion polarization focuses on identifying separate or orthogonal beliefs from media traces, thereby missing points of agreement among different communities. This paper develops a new class of Non-negative Matrix Factorization (NMF) algorithms that allow identification of both agreement and disagreement points when beliefs of different communities partially overlap. Specifically, we propose a novel Belief Structured Matrix Factorization algorithm (BSMF) to identify partially overlapping beliefs in polarized public social media. BSMF is totally unsupervised and considers three types of information: (i) who posted which opinion, (ii) keyword-level message similarity, and (iii) empirically observed social dependency graphs (e.g., retweet graphs), to improve belief separation. In the space of unsupervised belief separation algorithms, the emphasis was mostly given to the problem of identifying disjoint (e.g., conflicting) beliefs. The case when individuals with different beliefs agree on some subset of points was less explored. We observe that social beliefs overlap even in polarized scenarios. Our proposed unsupervised algorithm captures both the latent belief intersections and dissimilarities. We discuss the properties of the algorithm and conduct extensive experiments on both synthetic data and real-world datasets. The results show that our model outperforms all compared baselines by a great margin.

We describe an implementation of gradient boosting and neural guidance of saturation-style automated theorem provers that does not depend on consistent symbol names across problems. For the gradient-boosting guidance, we manually create abstracted features by considering arity-based encodings of formulas. For the neural guidance, we use symbol-independent graph neural networks and their embedding of the terms and clauses. The two methods are efficiently implemented in the E prover and its ENIGMA learning-guided framework and evaluated on the MPTP large-theory benchmark. Both methods are shown to achieve comparable real-time performance to state-of-the-art symbol-based methods.

As the thousands of high-flying dunks, alley oops and fadeaway jumper clips come out of NBA All-Star weekend, chances are the highlights were created by artificial intelligence. The league says they are using machine learning to create more highlights than ever before this All-Star weekend. Since 2014, the NBA has been employing and experimenting with technology from an Israeli company called WSC Sports to analyze key moments of each game and spit out some of the best highlights. As social media has emerged as an important destination to reach fans, the need for more and customized highlights has grown. This All-Star weekend, the software will automatically create multiple clips and content for every single player.

Wellington, Feb 12: New Zealand Police in a step to modernise its services unveiled its first Artificial intelligence (AI) officer named "Ella"on Wednesday. "Ella is a digital person that is powered by AI and uses real-time animation to emulate face-to face interactions," said Bush. Ella, which stands for'Electronic Lifelike Assistant', is part of two new digital kiosks the New Zealand Police has designed to help reduce queues in stations and to provide a modern way to connect with the public, NZ Herald reported. Ella will be stationed in the lobby of New Zealand Police's National Headquarters from next week assisting the concierge team and talking to visitors about Police topics such as the 105 nonemergency number and police vetting. Touch-screen electronic service points called Police Connect was also revealed on Wednesday, which will provide basic non-emergency services to the public 24 hours a day.

The police have unveiled their first AI officer, with hopes she'll soon be smiling and blinking out of screens in stations all around New Zealand. Ella, the artificial intelligence cop at the centre of the police's new digital services, was revealed at the police national headquarters in Wellington this morning. Ella, which stands for Electronic Lifelike Assistant, is part of two new digital kiosks police have designed to help reduce queues in stations and to provide a modern way to connect with the public. Designed as a mix of 26 different people, Ella is the brainchild of project manager Erin Greally, and will primarily be available only at the headquarters building in Molesworth St, where users can ask for information or be connected to whoever they're visiting. If the three-month pilot goes well, police hope to have Ella's friendly, CGI face spread across kiosks throughout the country.

Epoch gradient descent method (a.k.a. Epoch-GD) proposed by (Hazan and Kale, 2011) was deemed a breakthrough for stochastic strongly convex minimization, which achieves the optimal convergence rate of O(1/T) with T iterative updates for the objective gap. However, its extension to solving stochastic min-max problems with strong convexity and strong concavity still remains open, and it is still unclear whether a fast rate of O(1/T) for the duality gap is achievable for stochastic min-max optimization under strong convexity and strong concavity. Although some recent studies have proposed stochastic algorithms with fast convergence rates for min-max problems, they require additional assumptions about the problem, e.g., smoothness, bi-linear structure, etc. In this paper, we bridge this gap by providing a sharp analysis of epoch-wise stochastic gradient descent ascent method (referred to as Epoch-GDA) for solving strongly convex strongly concave (SCSC) min-max problems, without imposing any additional assumptions about smoothness or its structure. To the best of our knowledge, our result is the first one that shows Epoch-GDA can achieve the fast rate of O(1/T) for the duality gap of general SCSC min-max problems. We emphasize that such generalization of Epoch-GD for strongly convex minimization problems to Epoch-GDA for SCSC min-max problems is non-trivial and requires novel technical analysis. Moreover, we notice that the key lemma can be also used for proving the convergence of Epoch-GDA for weakly-convex strongly-concave min-max problems, leading to the best complexity as well without smoothness or other structural conditions.

We consider Model-Agnostic Meta-Learning (MAML) methods for Reinforcement Learning (RL) problems where the goal is to find a policy (using data from several tasks represented by Markov Decision Processes (MDPs)) that can be updated by one step of stochastic policy gradient for the realized MDP. In particular, using stochastic gradients in MAML update step is crucial for RL problems since computation of exact gradients requires access to a large number of possible trajectories. For this formulation, we propose a variant of the MAML method, named Stochastic Gradient Meta-Reinforcement Learning (SG-MRL), and study its convergence properties. We derive the iteration and sample complexity of SG-MRL to find an $\epsilon$-first-order stationary point, which, to the best of our knowledge, provides the first convergence guarantee for model-agnostic meta-reinforcement learning algorithms. We further show how our results extend to the case where more than one step of stochastic policy gradient method is used in the update during the test time.

This paper proposes an active learning system for sound event detection (SED). It aims at maximizing the accuracy of a learned SED model with limited annotation effort. The proposed system analyzes an initially unlabeled audio dataset, from which it selects sound segments for manual annotation. The candidate segments are generated based on a proposed change point detection approach, and the selection is based on the principle of mismatch-first farthest-traversal. During the training of SED models, recordings are used as training inputs, preserving the long-term context for annotated segments. The proposed system clearly outperforms reference methods in the two datasets used for evaluation (TUT Rare Sound 2017 and TAU Spatial Sound 2019). Training with recordings as context outperforms training with only annotated segments. Mismatch-first farthest-traversal outperforms reference sample selection methods based on random sampling and uncertainty sampling. Remarkably, the required annotation effort can be greatly reduced on the dataset where target sound events are rare: by annotating only 2% of the training data, the achieved SED performance is similar to annotating all the training data.

Recently, federated learning (FL) has drawn significant attention due to its capability of training a model over the network without knowing the client's private raw data. In this paper, we study the unsupervised clustering problem under the FL setting. By adopting a generalized matrix factorization model for clustering, we propose two novel (first-order) federated clustering (FedC) algorithms based on principles of model averaging and gradient sharing, respectively, and present their theoretical convergence conditions. We show that both algorithms have a O(1/T) convergence rate, where T is the total number of gradient evaluations per client, and the communication cost can be effectively reduced by controlling the local epoch length and allowing partial client participation within each communication round. Numerical experiments show that the FedC algorithm based on gradient sharing outperforms that based on model averaging, especially in scenarios with non-i.i.d. I. INTRODUCTION As one of the most fundamental data mining tasks, unsupervised clustering has a vast range of applications [1]. In view of the increasing volume of real-life data, distributed clustering methods that can process large-scale datasets in parallel computing environments have gained significant interests in the last decade [2], [3], [4]. However, recent emphasis on user privacy has called for new distributed schemes that can perform clustering without directly accessing the users' raw data.