Prediction tasks with high-dimensional nonorthogonal predictor sets pose a challenge for least squares based fitting procedures. A large and productive literature exists, discussing various regularized approaches to improving the out-of-sample robustness of parameter estimates. This paper proposes a novel cluster-based regularization -- the hierarchical feature regression (HFR) --, which mobilizes insights from the domains of machine learning and graph theory to estimate parameters along a supervised hierarchical representation of the predictor set, shrinking parameters towards group targets. The method is innovative in its ability to estimate optimal compositions of predictor groups, as well as the group targets endogenously. The HFR can be viewed as a supervised factor regression, with the strength of shrinkage governed by a penalty on the extent of idiosyncratic variation captured in the fitting process. The method demonstrates good predictive accuracy and versatility, outperforming a panel of benchmark regularized estimators across a diverse set of simulated regression tasks, including dense, sparse and grouped data generating processes. An application to the prediction of economic growth is used to illustrate the HFR's effectiveness in an empirical setting, with favorable comparisons to several frequentist and Bayesian alternatives.

Take it from U.S. Air Force Captain Kyle McAlpin when he says that scheduling C-17 aircraft crews is a headache. An artificial intelligence research flight commander for the Department of Air Force–MIT AI Accelerator Program, McAlpin is also an experienced C-17 pilot. "You could have a mission change and spend the next 12 hours of your life rebuilding a schedule that works," he says. It's a pain point for crew of 52 squadrons who operate C-17s, the military cargo aircraft that transport troops and supplies globally. This year, the Air Force marked 4 million flight hours for its C-17 fleet, which comprises 275 U.S. and allied aircraft.

We show that many machine-learning algorithms are specific instances of a single algorithm called the Bayesian learning rule. The rule, derived from Bayesian principles, yields a wide-range of algorithms from fields such as optimization, deep learning, and graphical models. This includes classical algorithms such as ridge regression, Newton's method, and Kalman filter, as well as modern deep-learning algorithms such as stochastic-gradient descent, RMSprop, and Dropout. The key idea in deriving such algorithms is to approximate the posterior using candidate distributions estimated by using natural gradients. Different candidate distributions result in different algorithms and further approximations to natural gradients give rise to variants of those algorithms. Our work not only unifies, generalizes, and improves existing algorithms, but also helps us design new ones.

Given a set of cities with certain locations, the Traveling Salesman Problem (TSP) is to find the shortest Hamiltonian route, along which a salesman travels from a city to visit all the cities exactly once and finally returns to the starting city. The TSP is one of the most famous and well-studied NP-hard combinatorial optimization problems, which is very easy to understand but very difficult to solve optimally or near-optimally. Over the years, TSP has become a touchstone for the algorithm design. Typical methods for solving the TSP are mainly exact algorithms, approximation algorithms and heuristics. The exact algorithms may be prohibitive for large instances and the approximation algorithms may suffer from weak optimal guarantees or empirical performance (Khalil et al. 2017). Heuristics are known to be the most efficient and effective approaches for solving the TSP.

Many modern machine learning applications come with complex and nuanced design goals such as minimizing the worst-case error, satisfying a given precision or recall target, or enforcing group-fairness constraints. Popular techniques for optimizing such non-decomposable objectives reduce the problem into a sequence of cost-sensitive learning tasks, each of which is then solved by re-weighting the training loss with example-specific costs. We point out that the standard approach of re-weighting the loss to incorporate label costs can produce unsatisfactory results when used to train over-parameterized models. As a remedy, we propose new cost-sensitive losses that extend the classical idea of logit adjustment to handle more general cost matrices. Our losses are calibrated, and can be further improved with distilled labels from a teacher model. Through experiments on benchmark image datasets, we showcase the effectiveness of our approach in training ResNet models with common robust and constrained optimization objectives.

Gradient descent is an optimization algorithm that follows the negative gradient of an objective function in order to locate the minimum of the function. A limitation of gradient descent is that it uses the same step size (learning rate) for each input variable. This can be a problem on objective functions that have different amounts of curvature in different dimensions, and in turn, may require a different sized step to a new point. Adaptive Gradients, or AdaGrad for short, is an extension of the gradient descent optimization algorithm that allows the step size in each dimension used by the optimization algorithm to be automatically adapted based on the gradients seen for the variable (partial derivatives) seen over the course of the search. In this tutorial, you will discover how to develop the gradient descent with adaptive gradients optimization algorithm from scratch.

Some real problems require the evaluation of expensive and noisy objective functions. Moreover, the analytical expression of these objective functions may be unknown. These functions are known as black-boxes, for example, estimating the generalization error of a machine learning algorithm and computing its prediction time in terms of its hyper-parameters. Multi-objective Bayesian optimization (MOBO) is a set of methods that has been successfully applied for the simultaneous optimization of black-boxes. Concretely, BO methods rely on a probabilistic model of the objective functions, typically a Gaussian process. This model generates a predictive distribution of the objectives. However, MOBO methods have problems when the number of objectives in a multi-objective optimization problem are 3 or more, which is the many objective setting. In particular, the BO process is more costly as more objectives are considered, computing the quality of the solution via the hyper-volume is also more costly and, most importantly, we have to evaluate every objective function, wasting expensive computational, economic or other resources. However, as more objectives are involved in the optimization problem, it is highly probable that some of them are redundant and not add information about the problem solution. A measure that represents how similar are GP predictive distributions is proposed. We also propose a many objective Bayesian optimization algorithm that uses this metric to determine whether two objectives are redundant. The algorithm stops evaluating one of them if the similarity is found, saving resources and not hurting the performance of the multi-objective BO algorithm. We show empirical evidence in a set of toy, synthetic, benchmark and real experiments that GPs predictive distributions of the effectiveness of the metric and the algorithm.

Gaussian processes with derivative information are useful in many settings where derivative information is available, including numerous Bayesian optimization and regression tasks that arise in the natural sciences. Incorporating derivative observations, however, comes with a dominating $O(N^3D^3)$ computational cost when training on $N$ points in $D$ input dimensions. This is intractable for even moderately sized problems. While recent work has addressed this intractability in the low-$D$ setting, the high-$N$, high-$D$ setting is still unexplored and of great value, particularly as machine learning problems increasingly become high dimensional. In this paper, we introduce methods to achieve fully scalable Gaussian process regression with derivatives using variational inference. Analogous to the use of inducing values to sparsify the labels of a training set, we introduce the concept of inducing directional derivatives to sparsify the partial derivative information of a training set. This enables us to construct a variational posterior that incorporates derivative information but whose size depends neither on the full dataset size $N$ nor the full dimensionality $D$. We demonstrate the full scalability of our approach on a variety of tasks, ranging from a high dimensional stellarator fusion regression task to training graph convolutional neural networks on Pubmed using Bayesian optimization. Surprisingly, we find that our approach can improve regression performance even in settings where only label data is available.

Gaussian process (GP) model based optimization is widely applied in simulation and machine learning. In general, it first estimates a GP model based on a few observations from the true response and then employs this model to guide the search, aiming to quickly locate the global optimum. Despite its successful applications, it has several limitations that may hinder its broader usage. First, building an accurate GP model can be difficult and computationally expensive, especially when the response function is multi-modal or varies significantly over the design space. Second, even with an appropriate model, the search process can be trapped in suboptimal regions before moving to the global optimum due to the excessive effort spent around the current best solution. In this work, we adopt the Additive Global and Local GP (AGLGP) model in the optimization framework. The model is rooted in the inducing-points-based GP sparse approximations and is combined with independent local models in different regions. With these properties, the AGLGP model is suitable for multi-modal responses with relatively large data sizes. Based on this AGLGP model, we propose a Combined Global and Local search for Optimization (CGLO) algorithm. It first divides the whole design space into disjoint local regions and identifies a promising region with the global model. Next, a local model in the selected region is fit to guide detailed search within this region. The algorithm then switches back to the global step when a good local solution is found. The global and local natures of CGLO enable it to enjoy the benefits of both global and local search to efficiently locate the global optimum.

Transfer learning methods aim to produce machine learning models that are trained on a given problem but perform well also on different new tasks. The interest in transfer learning comes from situations where large data sets can be used for solving a given training task but the data associated with new tasks are too small to train expressive models from scratch. The general transfer-learning strategy is to use the small available new data for adapting a large model that has been previously optimized on the training data. One option consists of keeping the structure of the pre-trained large model intact and fine-tuning its weights to solve the new task. When very few data points are available and the pre-trained network is large, however, customized regularization strategies are needed to mitigate the risk of over-fitting. Fine-tuning only a few parameters is a possible way out but can strongly limit the performance of the final model. Another option is to prune the pre-trained model to reduce its complexity, increase transferability, and prevent overfitting. Existing strategies, however, focus on optimized models and are unable to disentangle the network architecture from the attached weights. As a consequence, the pruned version of the original model can hardly be interpreted as a transferable new architecture and it is difficult to reuse it on new tasks.