Section: New Results
Participants : Imad Alawe, Yassine Hadjadj-Aoul, Corentin Hardy, Gerardo Rubino, Bruno Sericola, César Viho.
Distributed deep learning on edge-devices. A recently celebrated type of deep neural network is the Generative Adversarial Network (GAN). GANs are generators of samples from a distribution that has been learned; they are up to now centrally trained from local data on a single location. We question in  and in  the performance of training GANs using a spread dataset over a set of distributed machines, using a gossip approach shown to work on standard neural networks. This performance is compared to the federated learning distributed method, that has the drawback of sending model data to a server. We also propose a gossip variant, where GAN components are gossiped independently. Experiments are conducted with Tensorflow with up to 100 emulated machines, on the canonical MNIST dataset. The position of these papers is to provide a first evidence that gossip performances for GAN training are close to the ones of federated learning, while operating in a fully decentralized setup. Second, to highlight that for GANs, the distribution of data on machines is critical (i.e., i.i.d. or not). Third, to illustrate that the gossip variant, despite proposing data diversity to the learning phase, brings only marginal improvements over the classic gossip approach.
Machine learning acceleration. The number of connected devices is increasing with the emergence of new services and trends. This phenomenon is leading to a traffic growth over both the control and the data planes of the mobile core network. Therefore the 3GPP group has rethought the architecture of the New Generation Core (NGC) by defining its components as Virtualized Network Functions (VNF). However, scalability techniques should be envisioned in order to answer the needs, in term of resource provisioning, without degrading the Quality Of Service (QoS) already offered by hardware based core networks. Neural networks, and in particular deep learning, having shown their effectiveness in predicting time series , could be good candidates for predicting traffic evolution.
In , we proposed a novel solution to generalize neural networks while accelerating the learning process by using -mean clustering, and a Monte-Carlo method. We benchmarked multiple types of deep neural networks using real operator's data in order to compare their efficiency in predicting the upcoming network load for dynamic and proactive resource provisioning. The proposed solution allows obtaining very good predictions of the traffic evolution while reducing by 50% the time needed for the learning phase.
Machine Learning in Quality of Experience assessment. In a series of presentations we have disseminated the main ideas behind a new generation of Quality of Esperience assessing tools in preparation in the team. In the meetings  and , and also in the plenary , we described some of the key features of the tools we used in our PSQA project, the Random Neural Network of Erol Gelenbe, and the ideas we are following for extending some of their capabilities. The goal is to allow the user to evaluate with little additional cost, the sensitivities of the Quality of Experience with respect to specific metrics of interest, having in mind design applications, or improvement of existing systems. Another example is to invert the PSQA function providing a measure of the Quality of Experience as a function of several QoS and channel-based metrics, in order to define subsets of their joint state space where quality has a given property of interest (for instance, being good enough). In the plenary talk , we described other properties of these tools, and other directions being explored, such as the replacement of the subjective testing sessions leading to fully automatic tools, as well as to big data problems. In the keynote talk  we showed how to use our PSQA technology for classic performance evaluation works. The idea is that instead if targeting classic performance metrics such as a mean response time, or a loss rate (or dependability ones, the approach is the same), we can develop models that target the “ultimate goal”, the Quality of Experience itself. That is, instead of, say, providing a formula allowing to relate the loss rate of a system to the input data, we can obtain a (more complex) formula giving a numerical measure of the Quality of Experience as a function of the same data.