Models, code, and papers for "Gauri Joshi":

Communication-efficient SGD algorithms, which allow nodes to perform local updates and perform infrequent synchronization between them, are highly effective in improving the speed and scalability of distributed SGD. However, a rigorous convergence analysis and comparative study of different communication-reduction strategies remains a largely open problem. This paper presents a new framework called Cooperative SGD that subsumes existing communication-efficient SGD algorithms such as periodic-averaging, elastic-averaging and decentralized SGD. By analyzing Cooperative SGD, we provide novel convergence guarantees for existing algorithms. Moreover this framework enables us to design new communication-efficient SGD algorithms that strike the best balance between reducing communication overhead and achieving fast error convergence.

Large-scale machine learning training, in particular distributed stochastic gradient descent, needs to be robust to inherent system variability such as node straggling and random communication delays. This work considers a distributed training framework where each worker node is allowed to perform local model updates and the resulting models are averaged periodically. We analyze the true speed of error convergence with respect to wall-clock time (instead of the number of iterations), and analyze how it is affected by the frequency of averaging. The main contribution is the design of AdaComm, an adaptive communication strategy that starts with infrequent averaging to save communication delay and improve convergence speed, and then increases the communication frequency in order to achieve a low error floor. Rigorous experiments on training deep neural networks show that AdaComm can take $3 \times$ less time than fully synchronous SGD, and still reach the same final training loss.

We consider a correlated multi-armed bandit problem in which rewards of arms are correlated through a hidden parameter. Our approach exploits the correlation among arms to identify some arms as sub-optimal and pulls them only $\mathcal{O}(1)$ times. This results in significant reduction in cumulative regret, and in fact our algorithm achieves bounded (i.e., $\mathcal{O}(1)$) regret whenever possible; explicit conditions needed for bounded regret to be possible are also provided by analyzing regret lower bounds. We propose several variants of our approach that generalize classical bandit algorithms such as UCB, Thompson sampling, KL-UCB to the structured bandit setting, and empirically demonstrate their superiority via simulations.

We consider a novel multi-armed bandit framework where the rewards obtained by pulling the arms are functions of a common latent random variable. The correlation between arms due to the common random source can be used to design a generalized upper-confidence-bound (UCB) algorithm that identifies certain arms as $non-competitive$, and avoids exploring them. As a result, we reduce a $K$-armed bandit problem to a $C+1$-armed problem, where $C+1$ includes the best arm and $C$ $competitive$ arms. Our regret analysis shows that the competitive arms need to be pulled $\mathcal{O}(\log T)$ times, while the non-competitive arms are pulled only $\mathcal{O}(1)$ times. As a result, there are regimes where our algorithm achieves a $\mathcal{O}(1)$ regret as opposed to the typical logarithmic regret scaling of multi-armed bandit algorithms. We also evaluate lower bounds on the expected regret and prove that our correlated-UCB algorithm is order-wise optimal.

This paper studies the problem of {\em learning} the probability distribution $P_X$ of a discrete random variable $X$ using indirect and sequential samples. At each time step, we choose one of the possible $K$ functions, $g_1, \ldots, g_K$ and observe the corresponding sample $g_i(X)$. The goal is to estimate the probability distribution of $X$ by using a minimum number of such sequential samples. This problem has several real-world applications including inference under non-precise information and privacy-preserving statistical estimation. We establish necessary and sufficient conditions on the functions $g_1, \ldots, g_K$ under which asymptotically consistent estimation is possible. We also derive lower bounds on the estimation error as a function of total samples and show that it is order-wise achievable. Leveraging these results, we propose an iterative algorithm that i) chooses the function to observe at each step based on past observations; and ii) combines the obtained samples to estimate $p_X$. The performance of this algorithm is investigated numerically under various scenarios, and shown to outperform baseline approaches.

We consider a multi-armed bandit framework where the rewards obtained by pulling different arms are correlated. The correlation information is captured in terms of \textit{pseudo-rewards}, which are bounds on the rewards on the other arm given a reward realization and can capture many general correlation structures. We leverage these pseudo-rewards to design a novel approach that extends any classical bandit algorithm to the correlated multi-armed bandit setting studied in the framework. In each round, our proposed C-Bandit algorithm identifies some arms as empirically non-competitive, and avoids exploring them for that round. Through a unified regret analysis of the proposed C-Bandit algorithm, we show that C-UCB and C-TS (the correlated bandit versions of Upper-confidence-bound and Thompson sampling) pull certain arms called non-competitive arms, only O(1) times. As a result, we effectively reduce a $K$-armed bandit problem to a $C+1$-armed bandit problem, where $C$ is the number of competitive arms, as only $C$ sub-optimal arms are pulled O(log T) times. In many practical scenarios, $C$ can be zero due to which our proposed C-Bandit algorithms achieve bounded regret. In the special case where rewards are correlated through a latent random variable $X$, we give a regret lower bound that shows that bounded regret is possible only when $C = 0$. In addition to simulations, we validate the proposed algorithms via experiments on two real-world recommendation datasets, movielens and goodreads, and show that C-UCB and C-TS significantly outperform classical bandit algorithms.

Distributed Stochastic Gradient Descent (SGD) when run in a synchronous manner, suffers from delays in waiting for the slowest learners (stragglers). Asynchronous methods can alleviate stragglers, but cause gradient staleness that can adversely affect convergence. In this work we present a novel theoretical characterization of the speed-up offered by asynchronous methods by analyzing the trade-off between the error in the trained model and the actual training runtime (wallclock time). The novelty in our work is that our runtime analysis considers random straggler delays, which helps us design and compare distributed SGD algorithms that strike a balance between stragglers and staleness. We also present a new convergence analysis of asynchronous SGD variants without bounded or exponential delay assumptions, and a novel learning rate schedule to compensate for gradient staleness.

The trade-off between convergence error and communication delays in decentralized stochastic gradient descent~(SGD) is dictated by the sparsity of the inter-worker communication graph. In this paper, we propose MATCHA, a decentralized SGD method where we use matching decomposition sampling of the base graph to parallelize inter-worker information exchange so as to significantly reduce communication delay. At the same time, under standard assumptions for any general topology, in spite of the significant reduction of the communication delay, MATCHA maintains the same convergence rate as that of the state-of-the-art in terms of epochs. Experiments on a suite of datasets and deep neural networks validate the theoretical analysis and demonstrate the effectiveness of the proposed scheme as far as reducing communication delays is concerned.

Gaussian Processes (GPs) with an appropriate kernel are known to provide accurate predictions and uncertainty estimates even with very small amounts of labeled data. However, GPs are generally unable to learn a good representation that can encode intricate structures in high dimensional data. The representation power of GPs depends heavily on kernel functions used to quantify the similarity between data points. Traditional GP kernels are not very effective at capturing similarity between high dimensional data points, while methods that use deep neural networks to learn a kernel are not sample-efficient. To overcome these drawbacks, we propose deep probabilistic kernels which use a probabilistic neural network to map high-dimensional data to a probability distribution in a low dimensional subspace, and leverage the rich work on kernels between distributions to capture the similarity between these distributions. Experiments on a variety of datasets show that building a GP using this covariance kernel solves the conflicting problems of representation learning and sample efficiency. Our model can be extended beyond GPs to other small-data paradigms such as few-shot classification where we show competitive performance with state-of-the-art models on the mini-Imagenet dataset.

This paper introduces Selective-Backprop, a technique that accelerates the training of deep neural networks (DNNs) by prioritizing examples with high loss at each iteration. Selective-Backprop uses the output of a training example's forward pass to decide whether to use that example to compute gradients and update parameters, or to skip immediately to the next example. By reducing the number of computationally-expensive backpropagation steps performed, Selective-Backprop accelerates training. Evaluation on CIFAR10, CIFAR100, and SVHN, across a variety of modern image models, shows that Selective-Backprop converges to target error rates up to 3.5x faster than with standard SGD and between 1.02--1.8x faster than a state-of-the-art importance sampling approach. Further acceleration of 26% can be achieved by using stale forward pass results for selection, thus also skipping forward passes of low priority examples.

Federated learning (FL) is a machine learning setting where many clients (e.g. mobile devices or whole organizations) collaboratively train a model under the orchestration of a central server (e.g. service provider), while keeping the training data decentralized. FL embodies the principles of focused data collection and minimization, and can mitigate many of the systemic privacy risks and costs resulting from traditional, centralized machine learning and data science approaches. Motivated by the explosive growth in FL research, this paper discusses recent advances and presents an extensive collection of open problems and challenges.

Machine learning (ML) techniques are enjoying rapidly increasing adoption. However, designing and implementing the systems that support ML models in real-world deployments remains a significant obstacle, in large part due to the radically different development and deployment profile of modern ML methods, and the range of practical concerns that come with broader adoption. We propose to foster a new systems machine learning research community at the intersection of the traditional systems and ML communities, focused on topics such as hardware systems for ML, software systems for ML, and ML optimized for metrics beyond predictive accuracy. To do this, we describe a new conference, SysML, that explicitly targets research at the intersection of systems and machine learning with a program committee split evenly between experts in systems and ML, and an explicit focus on topics at the intersection of the two.