SGD with Momentum (SGDM) is widely used for large scale optimization of machine learning problems. Yet, the theoretical understanding of this algorithm is not complete. In fact, even the most recent results require changes to the algorithm like an averaging scheme and a projection onto a bounded domain, which are never used in practice. Also, no lower bound is known for SGDM. In this paper, we prove for the first time that for any constant momentum factor, there exists a Lipschitz and convex function for which the last iterate of SGDM suffers from an error $\Omega(\frac{\log T}{\sqrt{T}})$ after $T$ steps. Based on this fact, we study a new class of (both adaptive and non-adaptive) Follow-The-Regularized-Leader-based SGDM algorithms with \emph{increasing momentum} and \emph{shrinking updates}. For these algorithms, we show that the last iterate has optimal convergence $O (\frac{1}{\sqrt{T}})$ for unconstrained convex optimization problems. Further, we show that in the interpolation setting with convex and smooth functions, our new SGDM algorithm automatically converges at a rate of $O(\frac{\log T}{T})$. Empirical results are shown as well.

Matching markets are often organized in a multi-stage and decentralized manner. Moreover, participants in real-world matching markets often have uncertain preferences. This article develops a framework for learning optimal strategies in such settings, based on a nonparametric statistical approach and variational analysis. We propose an efficient algorithm, built upon concepts of "lower uncertainty bound" and "calibrated decentralized matching," for maximizing the participants' expected payoff. We show that there exists a welfare-versus-fairness trade-off that is characterized by the uncertainty level of acceptance. Participants will strategically act in favor of a low uncertainty level to reduce competition and increase expected payoff. We study signaling mechanisms that help to clear the congestion in such decentralized markets and find that the effects of signaling are heterogeneous, showing a dependence on the participants and matching stages. We prove that participants can be better off with multi-stage matching compared to single-stage matching. The deferred acceptance procedure assumes no limit on the number of stages and attains efficiency and fairness but may make some participants worse off than multi-stage matching. We demonstrate aspects of the theoretical predictions through simulations and an experiment using real data from college admissions.

In Apprenticeship Learning (AL), we are given a Markov Decision Process (MDP) without access to the cost function. Instead, we observe trajectories sampled by an expert that acts according to some policy. The goal is to find a policy that matches the expert's performance on some predefined set of cost functions. We introduce an online variant of AL (Online Apprenticeship Learning; OAL), where the agent is expected to perform comparably to the expert while interacting with the environment. We show that the OAL problem can be effectively solved by combining two mirror descent based no-regret algorithms: one for policy optimization and another for learning the worst case cost. To this end, we derive a convergent algorithm with $O(\sqrt{K})$ regret, where $K$ is the number of interactions with the MDP, and an additional linear error term that depends on the amount of expert trajectories available. Importantly, our algorithm avoids the need to solve an MDP at each iteration, making it more practical compared to prior AL methods. Finally, we implement a deep variant of our algorithm which shares some similarities to GAIL \cite{ho2016generative}, but where the discriminator is replaced with the costs learned by the OAL problem. Our simulations demonstrate our theoretically grounded approach outperforms the baselines.

We present a novel interactive learning protocol that enables training request-fulfilling agents by verbally describing their activities. Our protocol gives rise to a new family of interactive learning algorithms that offer complementary advantages against traditional algorithms like imitation learning (IL) and reinforcement learning (RL). We develop an algorithm that practically implements this protocol and employ it to train agents in two challenging request-fulfilling problems using purely language-description feedback. Empirical results demonstrate the strengths of our algorithm: compared to RL baselines, it is more sample-efficient; compared to IL baselines, it achieves competitive success rates while not requiring feedback providers to have agent-specific expertise. We also provide theoretical guarantees of the algorithm under certain assumptions on the teacher and the environment.

The test loss of well-trained neural networks often follows precise power-law scaling relations with either the size of the training dataset or the number of parameters in the network. We propose a theory that explains and connects these scaling laws. We identify variance-limited and resolution-limited scaling behavior for both dataset and model size, for a total of four scaling regimes. The variance-limited scaling follows simply from the existence of a well-behaved infinite data or infinite width limit, while the resolution-limited regime can be explained by positing that models are effectively resolving a smooth data manifold. In the large width limit, this can be equivalently obtained from the spectrum of certain kernels, and we present evidence that large width and large dataset resolution-limited scaling exponents are related by a duality. We exhibit all four scaling regimes in the controlled setting of large random feature and pretrained models and test the predictions empirically on a range of standard architectures and datasets. We also observe several empirical relationships between datasets and scaling exponents: super-classing image tasks does not change exponents, while changing input distribution (via changing datasets or adding noise) has a strong effect. We further explore the effect of architecture aspect ratio on scaling exponents.

We provide a new adaptive method for online convex optimization, MetaGrad, that is robust to general convex losses but achieves faster rates for a broad class of special functions, including exp-concave and strongly convex functions, but also various types of stochastic and non-stochastic functions without any curvature. We prove this by drawing a connection to the Bernstein condition, which is known to imply fast rates in offline statistical learning. MetaGrad further adapts automatically to the size of the gradients. Its main feature is that it simultaneously considers multiple learning rates, which are weighted directly proportional to their empirical performance on the data using a new meta-algorithm. We provide three versions of MetaGrad. The full matrix version maintains a full covariance matrix and is applicable to learning tasks for which we can afford update time quadratic in the dimension. The other two versions provide speed-ups for high-dimensional learning tasks with an update time that is linear in the dimension: one is based on sketching, the other on running a separate copy of the basic algorithm per coordinate. We evaluate all versions of MetaGrad on benchmark online classification and regression tasks, on which they consistently outperform both online gradient descent and AdaGrad.

When engineers train deep learning models, they are very much "flying blind". Commonly used approaches for real-time training diagnostics, such as monitoring the train/test loss, are limited. Assessing a network's training process solely through these performance indicators is akin to debugging software without access to internal states through a debugger. To address this, we present Cockpit, a collection of instruments that enable a closer look into the inner workings of a learning machine, and a more informative and meaningful status report for practitioners. It facilitates the identification of learning phases and failure modes, like ill-chosen hyperparameters. These instruments leverage novel higher-order information about the gradient distribution and curvature, which has only recently become efficiently accessible. We believe that such a debugging tool, which we open-source for PyTorch, represents an important step to improve troubleshooting the training process, reveal new insights, and help develop novel methods and heuristics.

We study a model selection problem in the linear bandit setting, where the learner must adapt to the dimension of the optimal hypothesis class on the fly and balance exploration and exploitation. More specifically, we assume a sequence of nested linear hypothesis classes with dimensions $d_1 < d_2 < \dots$, and the goal is to automatically adapt to the smallest hypothesis class that contains the true linear model. Although previous papers provide various guarantees for this model selection problem, the analysis therein either works in favorable cases when one can cheaply conduct statistical testing to locate the right hypothesis class or is based on the idea of "corralling" multiple base algorithms which often performs relatively poorly in practice. These works also mainly focus on upper bounding the regret. In this paper, we first establish a lower bound showing that, even with a fixed action set, adaptation to the unknown intrinsic dimension $d_\star$ comes at a cost: there is no algorithm that can achieve the regret bound $\widetilde{O}(\sqrt{d_\star T})$ simultaneously for all values of $d_\star$. We also bring new ideas, i.e., constructing virtual mixture-arms to effectively summarize useful information, into the model selection problem in linear bandits. Under a mild assumption on the action set, we design a Pareto optimal algorithm with guarantees matching the rate in the lower bound. Experimental results confirm our theoretical results and show advantages of our algorithm compared to prior work.

We address the problem of teaching a deep reinforcement learning (RL) agent to follow instructions in multi-task environments. We employ a well-known formal language -- linear temporal logic (LTL) -- to specify instructions, using a domain-specific vocabulary. We propose a novel approach to learning that exploits the compositional syntax and the semantics of LTL, enabling our RL agent to learn task-conditioned policies that generalize to new instructions, not observed during training. The expressive power of LTL supports the specification of a diversity of complex temporally extended behaviours that include conditionals and alternative realizations. Experiments on discrete and continuous domains demonstrate the strength of our approach in learning to solve (unseen) tasks, given LTL instructions.

Distillation is the technique of training a "student" model based on examples that are labeled by a separate "teacher" model, which itself is trained on a labeled dataset. The most common explanations for why distillation "works" are predicated on the assumption that student is provided with \emph{soft} labels, \eg probabilities or confidences, from the teacher model. In this work, we show, that, even when the teacher model is highly overparameterized, and provides \emph{hard} labels, using a very large held-out unlabeled dataset to train the student model can result in a model that outperforms more "traditional" approaches. Our explanation for this phenomenon is based on recent work on "double descent". It has been observed that, once a model's complexity roughly exceeds the amount required to memorize the training data, increasing the complexity \emph{further} can, counterintuitively, result in \emph{better} generalization. Researchers have identified several settings in which it takes place, while others have made various attempts to explain it (thus far, with only partial success). In contrast, we avoid these questions, and instead seek to \emph{exploit} this phenomenon by demonstrating that a highly-overparameterized teacher can avoid overfitting via double descent, while a student trained on a larger independent dataset labeled by this teacher will avoid overfitting due to the size of its training set.