Recent advances in neural architecture search (NAS) demand tremendous computational resources. This makes it difficult to reproduce experiments and imposes a barrier-to-entry to researchers without access to large-scale computation. We aim to ameliorate these problems by introducing NAS-Bench-101, the first public architecture dataset for NAS research. To build NAS-Bench-101, we carefully constructed a compact, yet expressive, search space, exploiting graph isomorphisms to identify 423k unique convolutional architectures. We trained and evaluated all of these architectures multiple times on CIFAR-10 and compiled the results into a large dataset. All together, NAS-Bench-101 contains the metrics of over 5 million models, the largest dataset of its kind thus far. This allows researchers to evaluate the quality of a diverse range of models in milliseconds by querying the pre-computed dataset. We demonstrate its utility by analyzing the dataset as a whole and by benchmarking a range of architecture optimization algorithms.

A novel machine learning optimization process coined Restrictive Federated Model Selection (RFMS) is proposed under the scenario, for example, when data from healthcare units can not leave the site it is situated on and it is forbidden to carry out training algorithms on remote data sites due to either technical or privacy and trust concerns. To carry out a clinical research under this scenario, an analyst could train a machine learning model only on local data site, but it is still possible to execute a statistical query at a certain cost in the form of sending a machine learning model to some of the remote data sites and get the performance measures as feedback, maybe due to prediction being usually much cheaper. Compared to federated learning, which is optimizing the model parameters directly by carrying out training across all data sites, RFMS trains model parameters only on one local data site but optimizes hyper-parameters across other data sites jointly since hyper-parameters play an important role in machine learning performance. The aim is to get a Pareto optimal model with respective to both local and remote unseen prediction losses, which could generalize well across data sites. In this work, we specifically consider high dimensional data with shifted distributions over data sites. As an initial investigation, Bayesian Optimization especially multi-objective Bayesian Optimization is used to guide an adaptive hyper-parameter optimization process to select models under the RFMS scenario. Empirical results show that solely using the local data site to tune hyper-parameters generalizes poorly across data sites, compared to methods that utilize the local and remote performances. Furthermore, in terms of dominated hypervolumes, multi-objective Bayesian Optimization algorithms show increased performance across multiple data sites among other candidates.

In optimization, the natural gradient method is well-known for likelihood maximization. The method uses the Kullback-Leibler divergence, corresponding infinitesimally to the Fisher-Rao metric, which is pulled back to the parameter space of a family of probability distributions. This way, gradients with respect to the parameters respect the Fisher-Rao geometry of the space of distributions, which might differ vastly from the standard Euclidean geometry of the parameter space, often leading to faster convergence. However, when minimizing an arbitrary similarity measure between distributions, it is generally unclear which metric to use. We provide a general framework that, given a similarity measure, derives a metric for the natural gradient. We then discuss connections between the natural gradient method and multiple other optimization techniques in the literature. Finally, we provide computations of the formal natural gradient to show overlap with well-known cases and to compute natural gradients in novel frameworks.

A central capability of intelligent systems is the ability to continuously build upon previous experiences to speed up and enhance learning of new tasks. Two distinct research paradigms have studied this question. Meta-learning views this problem as learning a prior over model parameters that is amenable for fast adaptation on a new task, but typically assumes the set of tasks are available together as a batch. In contrast, online (regret based) learning considers a sequential setting in which problems are revealed one after the other, but conventionally train only a single model without any task-specific adaptation. This work introduces an online meta-learning setting, which merges ideas from both the aforementioned paradigms to better capture the spirit and practice of continual lifelong learning. We propose the follow the meta leader algorithm which extends the MAML algorithm to this setting. Theoretically, this work provides an $\mathcal{O}(\log T)$ regret guarantee with only one additional higher order smoothness assumption in comparison to the standard online setting. Our experimental evaluation on three different large-scale tasks suggest that the proposed algorithm significantly outperforms alternatives based on traditional online learning approaches.

Comparing different algorithms is hard. For almost any pair of algorithms and measure of algorithm performance like running time or solution quality, each algorithm will perform better than the other on some inputs.a For example, the insertion sort algorithm is faster than merge sort on already-sorted arrays but slower on many other inputs. When two algorithms have incomparable performance, how can we deem one of them "better than" the other? Worst-case analysis is a specific modeling choice in the analysis of algorithms, where the overall performance of an algorithm is summarized by its worst performance on any input of a given size. The "better" algorithm is then the one with superior worst-case performance. Merge sort, with its worst-case asymptotic running time of Θ(n log n) for arrays of length n, is better in this sense than insertion sort, which has a worst-case running time of Θ(n2). While crude, worst-case analysis can be tremendously useful, and it is the dominant paradigm for algorithm analysis in theoretical computer science. A good worst-case guarantee is the best-case scenario for an algorithm, certifying its general-purpose utility and absolving its users from understanding which inputs are relevant to their applications. Remarkably, for many fundamental computational problems, there are algorithms with excellent worst-case performance guarantees. The lion's share of an undergraduate algorithms course comprises algorithms that run in linear or near-linear time in the worst case.

We derive an optimal policy for adaptively restarting a randomized algorithm, based on observed features of the run-so-far, so as to minimize the expected time required for the algorithm to successfully terminate. Given a suitable Bayesian prior, this result can be used to select the optimal black-box optimization algorithm from among a large family of algorithms that includes random search, Successive Halving, and Hyperband. On CIFAR-10 and ImageNet hyperparameter tuning problems, the proposed policies offer up to a factor of 13 improvement over random search in terms of expected time to reach a given target accuracy, and up to a factor of 3 improvement over a baseline adaptive policy that terminates a run whenever its accuracy is below-median.

A rapidly growing area of work has studied the existence of adversarial examples, datapoints which have been perturbed to fool a classifier, but the vast majority of these works have focused primarily on threat models defined by $\ell_p$ norm-bounded perturbations. In this paper, we propose a new threat model for adversarial attacks based on the Wasserstein distance. In the image classification setting, such distances measure the cost of moving pixel mass, which naturally cover "standard" image manipulations such as scaling, rotation, translation, and distortion (and can potentially be applied to other settings as well). To generate Wasserstein adversarial examples, we develop a procedure for projecting onto the Wasserstein ball, based upon a modified version of the Sinkhorn iteration. The resulting algorithm can successfully attack image classification models, bringing traditional CIFAR10 models down to 3% accuracy within a Wasserstein ball with radius 0.1 (i.e., moving 10% of the image mass 1 pixel), and we demonstrate that PGD-based adversarial training can improve this adversarial accuracy to 76%. In total, this work opens up a new direction of study in adversarial robustness, more formally considering convex metrics that accurately capture the invariances that we typically believe should exist in classifiers. Code for all experiments in the paper is available at https://github.com/locuslab/projected_sinkhorn.

Data science relies on pipelines that are organized in the form of interdependent computational steps. Each step consists of various candidate algorithms that maybe used for performing a particular function. Each algorithm consists of several hyperparameters. Algorithms and hyperparameters must be optimized as a whole to produce the best performance. Typical machine learning pipelines consist of complex algorithms in each of the steps. Not only is the selection process combinatorial, but it is also important to interpret and understand the pipelines. We propose a method to quantify the importance of different components in the pipeline, by computing an error contribution relative to an agnostic choice of computational steps, algorithms and hyperparameters. We also propose a methodology to quantify the propagation of error from individual components of the pipeline with the help of a naive set of benchmark algorithms not involved in the pipeline. We demonstrate our methodology on image classification pipelines. The agnostic and naive methodologies quantify the error contribution and propagation respectively from the computational steps, algorithms and hyperparameters in the image classification pipeline. We show that algorithm selection and hyperparameter optimization methods like grid search, random search and Bayesian optimization can be used to quantify the error contribution and propagation, and that random search is able to quantify them more accurately than Bayesian optimization. This methodology can be used by domain experts to understand machine learning and data analysis pipelines in terms of their individual components, which can help in prioritizing different components of the pipeline.

This paper introduces an elasticity reconstruction method based on local displacement observations of elastic bodies. Sparse reconstruction theory is applied to formulate the underdetermined inverse problems of elasticity reconstruction including unobserved areas. An online local clustering scheme called a superelement is proposed to reduce the number of dimensions of the optimization parameters. Alternating the optimization of element boundaries and elasticity parameters enables the elasticity distribution to be estimated with a higher spatial resolution. The simulation experiments show that elasticity distribution is reconstructed based on observations of approximately 10% of the total body. The estimation error was improved when considering the sparseness of the elasticity distribution.

The min-max problem, also known as the saddle point problem, is a class of optimization problems in which we minimize and maximize two subsets of variables simultaneously. This class of problems can be used to formulate a wide range of signal processing and communication (SPCOM) problems. Despite its popularity, existing theory for this class has been mainly developed for problems with certain special convex-concave structure. Therefore, it cannot be used to guide the algorithm design for many interesting problems in SPCOM, where some kind of non-convexity often arises. In this work, we consider a general block-wise one-sided non-convex min-max problem, in which the minimization problem consists of multiple blocks and is non-convex, while the maximization problem is (strongly) concave. We propose a class of simple algorithms named Hybrid Block Successive Approximation (HiBSA), which alternatingly performs gradient descent-type steps for the minimization blocks and one gradient ascent-type step for the maximization problem. A key element in the proposed algorithm is the introduction of certain properly designed regularization and penalty terms, which are used to stabilize the algorithm and ensure convergence. For the first time, we show that such simple alternating min-max algorithms converge to first-order stationary solutions, with quantifiable global rates. To validate the efficiency of the proposed algorithms, we conduct numerical tests on a number of information processing and wireless communication problems, including the robust learning problem, the non-convex min-utility maximization problems, and certain wireless jamming problem arising in interfering channels.