arXiv subjects

Bruno Sericola

Publications and source records attributed to Bruno Sericola.

7 recordsLinked to original sources

arXiv2019

MD-GAN: Multi-Discriminator Generative Adversarial Networks for Distributed Datasets

A recent technical breakthrough in the domain of machine learning is the discovery and the multiple applications of Generative Adversarial Networks (GANs). Those generative models are computationally demanding, as a GAN is composed of two deep neural networks, and because it trains on large datasets. A GAN is generally trained on a single server. In this paper, we address the problem of distributing GANs so that they are able to train over datasets that are spread on multiple workers. MD-GAN is exposed as the first solution for this problem: we propose a novel learning procedure for GANs so that they fit this distributed setup. We then compare the performance of MD-GAN to an adapted version of Federated Learning to GANs, using the MNIST and CIFAR10 datasets. MD-GAN exhibits a reduction by a factor of two of the learning complexity on each worker node, while providing better performances than federated learning on both datasets. We finally discuss the practical implications of distributing GANs.

cs.LG↗

arXiv2017

Distributed deep learning on edge-devices: feasibility via adaptive compression

A large portion of data mining and analytic services use modern machine learning techniques, such as deep learning. The state-of-the-art results by deep learning come at the price of an intensive use of computing resources. The leading frameworks (e.g., TensorFlow) are executed on GPUs or on high-end servers in datacenters. On the other end, there is a proliferation of personal devices with possibly free CPU cycles; this can enable services to run in users' homes, embedding machine learning operations. In this paper, we ask the following question: Is distributed deep learning computation on WAN connected devices feasible, in spite of the traffic caused by learning tasks? We show that such a setup rises some important challenges, most notably the ingress traffic that the servers hosting the up-to-date model have to sustain. In order to reduce this stress, we propose adaComp, a novel algorithm for compressing worker updates to the model on the server. Applicable to stochastic gradient descent based approaches, it combines efficient gradient selection and learning rate modulation. We then experiment and measure the impact of compression, device heterogeneity and reliability on the accuracy of learned models, with an emulator platform that embeds TensorFlow into Linux containers. We report a reduction of the total amount of data sent by workers to the server by two order of magnitude (e.g., 191-fold reduction for a convolutional network on the MNIST dataset), when compared to a standard asynchronous stochastic gradient descent, while preserving model accuracy.

cs.LG↗

arXiv2015

Optimization results for a generalized coupon collector problem

We study in this paper a generalized coupon collector problem, which consists in analyzing the time needed to collect a given number of distinct coupons that are drawn from a set of coupons with an arbitrary probability distribution. We suppose that a special coupon called the null coupon can be drawn but never belongs to any collection. In this context, we prove that the almost uniform distribution, for which all the non-null coupons have the same drawing probability, is the distribution which stochastically minimizes the time needed to collect a fixed number of distinct coupons. Moreover, we show that in a given closed subset of probability distributions, the distribution with all its entries, but one, equal to the smallest possible value is the one, which stochastically maximizes the time needed to collect a fixed number of distinct coupons. An computer science application shows the utility of these results.

cs.DS↗

arXiv2014

Impatience in mobile networks and its application to data pricing

We consider in this paper an import Quality of Experience (QoE) indicator in mobile networks that is reneging of users due to impatience. We specifically consider a cell under heavy load conditions and compute the reneging probability by using a fluid limit analysis. By solving the fixed point equation, we obtain a new QoE perturbation metric quantifying the impact of reneging on the performance of the system. This metric is then used to devise a new pricing scheme accounting of reneging. We specifically propose several flavors of this pricing around the idea of having a flat rate for accessing the network and an elastic price related to the level of QoE perturbation induced by communications.

cs.NI↗

arXiv2014

New results on a generalized coupon collector problem using Markov chains

We study in this paper a generalized coupon collector problem, which consists in determining the distribution and the moments of the time needed to collect a given number of distinct coupons that are drawn from a set of coupons with an arbitrary probability distribution. We suppose that a special coupon called the null coupon can be drawn but never belongs to any collection. In this context, we obtain expressions of the distribution and the moments of this time. We also prove that the almost-uniform distribution, for which all the non-null coupons have the same drawing probability, is the distribution which minimizes the expected time to get a fixed subset of distinct coupons. This optimization result is extended to the complementary distribution of that time when the full collection is considered, proving by the way this well-known conjecture. Finally, we propose a new conjecture which expresses the fact that the almost-uniform distribution should minimize the complementary distribution of the time needed to get any fixed number of distinct coupons.

math.PR↗

arXiv2008

Ascending runs in dependent uniformly distributed random variables: Application to wireless networks

We analyze in this paper the longest increasing contiguous sequence or maximal ascending run of random variables with common uniform distribution but not independent. Their dependence is characterized by the fact that two successive random variables cannot take the same value. Using a Markov chain approach, we study the distribution of the maximal ascending run and we develop an algorithm to compute it. This problem comes from the analysis of several self-organizing protocols designed for large-scale wireless sensor networks, and we show how our results apply to this domain.

cs.DM↗

arXiv2008

Core Persistence in Peer-to-Peer Systems: Relating Size to Lifetime

Distributed systems are now both very large and highly dynamic. Peer to peer overlay networks have been proved efficient to cope with this new deal that traditional approaches can no longer accommodate. While the challenge of organizing peers in an overlay network has generated a lot of interest leading to a large number of solutions, maintaining critical data in such a network remains an open issue. In this paper, we are interested in defining the portion of nodes and frequency one has to probe, given the churn observed in the system, in order to achieve a given probability of maintaining the persistence of some critical data. More specifically, we provide a clear result relating the size and the frequency of the probing set along with its proof as well as an analysis of the way of leveraging such an information in a large scale dynamic distributed system.

cs.DC↗