In many situations, such as robot-learning, training experience is very expensive. One strategy for reducing the amount of training data needed for a new task is to learn some form of prior or bias using data from several related tasks. The objective of this process is to extract information that will substantially reduce the training-data requirements for a new task. This problem is a form of transfer learning, sometimes also called meta-learning or "learning to learn" [1, 2]. Previous approaches to meta-learning for robotics have focused on finding distributions over [3] or initial values of [4, 5] parameters, based on a set of "training tasks," that will enable a new "test task" to be learned with many fewer training examples. Our objective is similar, but rather than focusing on transferring information about parameter values, we focus on finding a reusable set of modules that can form components of a solution to a new task, possibly with a small amount of tuning. Modular approaches to learning have been very successful in structured tasks such as naturallanguage sentence interpretation [6], in which the input signal gives relatively direct information about a good structural decomposition of the problem. We wish to address problems that may benefit from a modular decomposition but do not provide any task-level input from which the structure of a solution may be derived. Nonetheless, we adopt a similar modular structure and parameteradaptation method for learning our reusable modules, but use a general-purpose simulated-annealing search strategy to find an appropriate structural decomposition for each new task.

Many modern machine learning models are trained to achieve zero or near-zero training error in order to obtain near-optimal (but non-zero) test error. This phenomenon of strong generalization performance for "overfitted" / interpolated classifiers appears to be ubiquitous in high-dimensional data, having been observed in deep networks, kernel machines, boosting and random forests. Their performance is robust even when the data contain large amounts of label noise. Very little theory is available to explain these observations. The vast majority of theoretical analyses of generalization allows for interpolation only when there is little or no label noise. This paper takes a step toward a theoretical foundation for interpolated classifiers by analyzing local interpolating schemes, including geometric simplicial interpolation algorithm and weighted $k$-nearest neighbor schemes. Consistency or near-consistency is proved for these schemes in classification and regression problems. These schemes have an inductive bias that benefits from higher dimension, a kind of "blessing of dimensionality". Finally, connections to kernel machines, random forests, and adversarial examples in the interpolated regime are discussed.

A long-standing problem in the theory of stochastic gradient descent (SGD) is to prove that its without-replacement version RandomShuffle converges faster than the usual with-replacement version. We present the first (to our knowledge) non-asymptotic solution to this problem, which shows that after a "reasonable" number of epochs RandomShuffle indeed converges faster than SGD. Specifically, we prove that under strong convexity and second-order smoothness, the sequence generated by RandomShuffle converges to the optimal solution at the rate O(1/T^2 + n^3/T^3), where n is the number of components in the objective, and T is the total number of iterations. This result shows that after a reasonable number of epochs RandomShuffle is strictly better than SGD (which converges as O(1/T)). The key step toward showing this better dependence on T is the introduction of n into the bound; and as our analysis will show, in general a dependence on n is unavoidable without further changes to the algorithm. We show that for sparse data RandomShuffle has the rate O(1/T^2), again strictly better than SGD. Furthermore, we discuss extensions to nonconvex gradient dominated functions, as well as non-strongly convex settings.

We consider the following conditional linear regression problem: the task is to identify both (i) a $k$-DNF condition $c$ and (ii) a linear rule $f$ such that the probability of $c$ is (approximately) at least some given bound $\mu$, and $f$ minimizes the $\ell_p$ loss of predicting the target $z$ in the distribution of examples conditioned on $c$. Thus, the task is to identify a portion of the distribution on which a linear rule can provide a good fit. Algorithms for this task are useful in cases where simple, learnable rules only accurately model portions of the distribution. The prior state-of-the-art for such algorithms could only guarantee finding a condition of probability $\Omega(\mu/n^k)$ when a condition of probability $\mu$ exists, and achieved an $O(n^k)$-approximation to the target loss, where $n$ is the number of Boolean attributes. Here, we give efficient algorithms for solving this task with a condition $c$ that nearly matches the probability of the ideal condition, while also improving the approximation to the target loss. We also give an algorithm for finding a $k$-DNF reference class for prediction at a given query point, that obtains a sparse regression fit that has loss within $O(n^k)$ of optimal among all sparse regression parameters and sufficiently large $k$-DNF reference classes containing the query point.

We provide tight finite-time convergence bounds for gradient descent and stochastic gradient descent on quadratic functions, when the gradients are delayed and reflect iterates from $\tau$ rounds ago. First, we show that without stochastic noise, delays strongly affect the attainable optimization error: In fact, the error can be as bad as non-delayed gradient descent ran on only $1/\tau$ of the gradients. In sharp contrast, we quantify how stochastic noise makes the effect of delays negligible, improving on previous work which only showed this phenomenon asymptotically or for much smaller delays. Also, in the context of distributed optimization, the results indicate that the performance of gradient descent with delays is competitive with synchronous approaches such as mini-batching. Our results are based on a novel technique for analyzing convergence of optimization algorithms using generating functions.

Budgeted Stochastic Gradient Descent (BSGD) is a state-of-the-art technique for training large-scale kernelized support vector machines. The budget constraint is maintained incrementally by merging two points whenever the pre-defined budget is exceeded. The process of finding suitable merge partners is costly; it can account for up to 45% of the total training time. In this paper we investigate computationally more efficient schemes that merge more than two points at once. We obtain significant speed-ups without sacrificing accuracy.

Support Vector Machines (SVMs) introduced by [5] are popular machine learning methods, in particular for binary classification. They are supported by learning-theoretical guarantees [14], and they exhibit excellent generalization performance in many applications in science and technology [1, 16, 29, 23, 22, 3, 18, 19, 10]. They belong to the family of kernel methods, applying a linear algorithm in a feature space defined implicitly by a kernel function. Training an SVM corresponds to solving a large-scale optimization problem, which can be cast into a quadratic program (QP). The primal problem can be solved directly with stochastic gradient descent (SGD) and accelerated variants [21, 8], while the dual QP is solved with subspace ascent, see [2] and references therein.

Limiting the model size of a kernel support vector machine to a pre-defined budget is a well-established technique that allows to scale SVM learning and prediction to large-scale data. Its core addition to simple stochastic gradient training is budget maintenance through merging of support vectors. This requires solving an inner optimization problem with an iterative method many times per gradient step. In this paper we replace the iterative procedure with a fast lookup. We manage to reduce the merging time by up to 65% and the total training time by 44% without any loss of accuracy.

We study in this paper the problem of jointly clustering and learning representations. As several previous studies have shown, learning representations that are both faithful to the data to be clustered and adapted to the clustering algorithm can lead to better clustering performance, all the more so that the two tasks are performed jointly. We propose here such an approach for $k$-Means clustering based on a continuous reparametrization of the objective function that leads to a truly joint solution. The behavior of our approach is illustrated on various datasets showing its efficacy in learning representations for objects while clustering them.

The problem of low-rank matrix estimation recently received a lot of attention due to challenging applications. A lot of work has been done on rank-penalized methods and convex relaxation, both on the theoretical and applied sides. However, only a few papers considered Bayesian estimation. In this paper, we review the different type of priors considered on matrices to favour low-rank. We also prove that the obtained Bayesian estimators, under suitable assumptions, enjoys the same optimality properties as the ones based on penalization.