978-1-7281-3627-1/19/$31.00 ©2019 IEEE Deploying Artificial Intelligence in the Wireless Infrastructure: the Challenges Ahead Miguel Ángel Vázquez Centre Tecnològic de Telecomunicacions de Catalunya Castelldefels, Spain mavazquez@cttc.es Christos Masouros University College London London, United Kingdom c.masouros@ucl.ac.uk Fisseha Mekuria Council for Scientific and Industrial Research Pretoria, South Africa FMekuria@csir.co.za Jean Paul Pallois Huawei Technologies Paris, France jp.pallois@huawei.com Tony Kenyon University College London London, United Kingdom a.kenyon@ucl.ac.uk Ana Pérez-Neira Centre Tecnològic de Telecomunicacions de Catalunya and Univeritat Politècnica de Catalunya Castelldefels, Spain aperez@cttc.es Merouane Debbah Huawei Research Centre Paris, France merouane.debbah@centralesupelec.fr Yansha Deng King’s College London London, United Kingdom yansha.deng@kcl.ac.uk Javan Erfanian Bell Mobility Ontario, Canada javan.erfanian@bell.ca Abstract— The adoption of artificial intelligence (AI) techniques entails a substantial change in the wireless ecosystem where data as well as their owners become crucial. As a result, the roll out of AI techniques in wireless systems raises a plethora of questions. In this context, we describe the challenges observed by the wireless stakeholders when deploying AI. Furthermore, we introduce the recent discussion in field of ethics that appear when managing wireless communications data. Keywords—Artificial Intelligence, Machine Learning, Wireless Infrastructure, I. INTRODUCTION Telecom operators see AI as almost the unavoidable technology enabling to maintain or even reduce the operational expenditures (OPEX) significantly while delivering higher quality-of-service (QoS) to the end-users. Communication service providers will be looking at AI for various reasons. On top of enhancing their business ability, the adoption of AI will be driven by the increased complexity of the network. Indeed, with the advent of 5G, networks will be handling more spectrum, more and varied bandwidths, additional radio technologies, dealing with lower latencies enabling to reach new business territories like on-line tactile applications while benefiting from always more computing power. AI has the power to change the networks of the future from reactive networks, to predictive and finally proactive networks. Insights from the deep learning AI systems using the huge amount of data generated by the complex wireless systems of the future, will be able to support effective utilization of spectrum & radio resources, self-optimization of network operations and insight based QoS provisioning, thereby benefitting the wide spread application of AI in telecoms. However, AI is not seen as the grail technology by everyone. A few factors, not necessarily negligible, may slow down the adoption of AI by the industry. The success of AI in wireless networks of the future therefore hinges, in developing solutions that will address these challenges starting from the design of the network layers and up to the service layers. In this paper, we describe those factors in the next Section. In addition, in Section III we describe the first issues raised by the stakeholders regarding the proper use of the wireless data. Section IV concludes. II. AI CHALLENGES The first reason may rely on the difficulty to fully assess the real benefits AI can bring in daily work despite the fact the market feels it is the way to go. Another reason is also a human, emotional reason: people in the industry may perceive AI as a potential threat to jobs. Right or wrong, this is a reality, AI proponents will have to deal with. But even if the AI advocates pass these barriers, the need to adapt existing network operations processes will be a heavy, cumbersome and tedious task and will need support from the top management in organizations adopting AI. An additional obstacle will be the expected transparency of AI. Indeed, delivering AI as a black box may be enough in some cases for basic actions (e.g. like how to optimize parameter settings during a roll-out phase) but would not be acceptable during certain or healing phases. Operators who could not be in the position to explain what happened and why it happened could be reliable to liabilities. Here in lies the next obstacle, which is the training data set accuracy and inclusiveness, and might result in a biased decision making which can be detrimental to the company and its customer base. This is why explainable AI will be sooner or later necessary. All these factors will need, undoubtedly, to be addressed from a successful adoption of AI by the market and the telecommunications industry. Importantly, in the dynamic and real time wireless transmission scenarios of the future, the training overheads for learning based AI solutions may pose a particular challenge. As an example, typical learning tools available have been developed for applications such as computer vision, speech recognition, or natural language processing, where training time and overheads are not a key factor. In a dynamic wireless environment where the transceivers will need to adopt to changing channels, hardware responses, or event dynamic link connectivity, the overheads of re-training