Section: New Results

Machine learning

Participants : Yassine Hadjadj-Aoul, Corentin Hardy, Quang Pham Tran Anh, Gerardo Rubino, Bruno Sericola, Imane Taibi, 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 [37] the performance of training GANs using a spread dataset over a set of distributed machines, following 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 the paper 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.

This work is a part of the thesis [14].

Deep reinforcement learning for network slicing. Recent achievements in Deep Reinforcement Learning (DRL) have shown the potential of these approaches to solve combinatorial optimization problems. However, the Deep Deterministic Policy Gradient algorithm (DDPG), which is one of the most effective techniques, is not suitable to deal with large-scale discrete action space, which is the case of the Virtual Network Function-Forwarding Graph (VNF-FG) placement. To deal with this problem, we propose several enhancements to improve DDPG efficiency [25][47]. The conventional DDPG generates only one action per iteration; thus, it slowly explores the action space especially in a large action space. Thus, we propose to enhance the exploration by considering multiple noisy actions. In order to avoid getting stuck at a local minimum, we propose to multiply the number of critic (for Q-value) neural networks [25]. In order to improve further the exploration, we propose in [47] an evolutionary algorithm to evolve these neural networks in order to discover better ones.

The techniques presented above are generic and can be applied to a variety of problems. To make them even more effective for network slicing problems, we have also proposed to combine them with a proposed First-Fit heuristic that allows for even more interesting results.

Machine learning for Indoor Outdoor detection. Detecting whether a mobile user is indoor or outdoor is an important issue which significantly impacts user behavior contextualization and mobile network resource management. In [59] we investigate hybrid/semi-supervised Deep Learning-based methods for detecting the environment of an active mobile phone user. They are based on both labeled and unlabeled large real radio data obtained from inside the network and from 3GPP signal measurements. We have empirically evaluated the effectiveness of the semi-supervised learning methods using new real-time radio data, with partial ground truth information, gathered massively from multiple typical and diversified locations (indoor and outdoor) of mobile users. We also presented an analysis of such schemes as compared to the existing supervised classification methods including SVM and Deep Learning [57].

Cognition of user behavior can be seen as an efficient tool for automation of future mobile networks. The work presented in [51] deals with the user behaviour modeling. The model includes the prediction of two main features related to mobile user context: the environment and the mobility. We investigate Deep Learning based methods for simultaneously detecting the environment and the mobility state. We empirically evaluate the effectiveness of the proposed techniques using real-time radio data, which has been massively gathered from multiple diversified situations of mobile users.

Predicting the future Perceived Quality level with PSQA. PSQA is a technology developed by Dionysos during a period of several years, whose aim is quantifying the Quality of Experience (more precisely, the Perceived Quality) of an application or service built on the Internet around the transport of audio or video-audio signals. The main properties of PSQA are the its accuracy (indistinguishable from a subjective testing session), the fact that it is fully automatic, with no reference, and able to operate in real time. PSQA is based on supervised learning (the tool learns from subjective testing panels); once trained and validated, it works with no human intervention. In the PSQA project we selected the Random Neural Network tool for the supervised learning associated tasks, after a comparison with the available techniques at the beginning of the project. In [31] we recall all these elements, including the numerical aspects on the optimization side of the learning process, and then, we focus in the current developments where the goal is to predict the Perceived Quality in the close future. This includes the description of the Reservoir Computing models for time series forecasting, and of a tool we proposed, called Echo State Queueing Network, which is a mix between Reservoir Computing and Random Neural Networks.