Post-Doctoral Research Visit – Distributed Machine Learning at the Network Edge

Type de contrat : CDD Contexte et atouts du poste The position is in the framework of dAIEDGE---A network of excellence for distributed, trustworthy, efficient and scalable AI at the Edge---funded by the European Union. The vision of the dAIEDGE Network of Excellence is to strengthen and support the development of the dynamic European edge AI ecosystem under the umbrella of the European AI Lighthouse and to sustain the advanced research and innovation of distributed AI at the edge as essential digital, enabling, and emerging technology in an extensive range of industrial sectors. The candidate will work with NEO Inria team (https://team.inria.fr/neo/) and COATI Inria team (https://team.inria.fr/coati/), and in particular with • Giovanni Neglia • Chuan Xu • Frédéric Giroire Mission confiée The Internet was conceived to enable computer resources’ time-sharing, but soon its main function became to deliver content to end users, but it is now called to play a new key role: to pervasively support machine learning (ML) operation both for model training and prediction serving. There are two aspects calling for Internet-wide deployment of ML systems. First, data---one key ingredient of ML success---is often generated by users and devices at the edge of the network. The classic ML operation in the cloud requires such data to be collected at a single computing facility where training occurs. Data aggregation can be very costly, or simply impossible because of capacity constraints, privacy issues, or ownership ones. These scenarios call for distributed learning systems, where computation moves, at least in part, to the data. For example, Google’s federated learning [mcmahan17,kairouz21] enables mobile phones, or other devices with limited computing capabilities, to collaboratively learn an ML model while keeping all training data locally. Distributed ML training is already a difficult task in a cluster setting. Indeed, optimization techniques, distributed systems, and ML models are a triad difficult to untangle: e.g., relaxed state consistency across computing nodes increases system throughput but may jeopardize convergence of the optimization algorithm or affect the final solution selected, leading to models with very different generalization capabilities [chen16]. Additional challenges arise when training moves to the Internet. First, the system potentially scales up to billions of devices, against at most thousands of GPUs to break ML training records in a cluster. Second, local datasets are highly heterogeneous with very different sizes and feature/label distributions. Third, devices may have very different hardware and connectivity. Fourth, communications are often unreliable (devices can be switched off at any time), slow (latencies are 2 orders of magnitude larger), and expensive for battery-constrained devices. Fifth, privacy concerns are often important and limit the operations that can be performed during training to avoid inadvertently disclosing sensible information. Finally, training is more vulnerable to malicious attacks. For all these reasons, federated learning (as ML training over the Internet is now usually called) has emerged in the last years as a specific research topic—well distinct for example from high-performance computing or cloud computing—at the intersection of machine learning, optimization, distributed systems, and networking. The second driver to distribute ML processes over the Internet is real-time inference. In fact, ML models are often trained for inference’s purposes, i.e., to make predictions on new data. Model predictions need then to be served to the final users. ML training is a computationally expensive operation and is the object of much research effort. Inference does not involve complex iterative algorithms and is therefore generally assumed to be easy, but it also presents fundamental challenges that are likely to become dominant as ML adoption increases [stoica17]. AI systems will be ubiquitously deployed and will need to make timely and safe decisions in unpredictable environments. In this case, inference must run in real-time, and predictions may need to be served at a very high rate. The big cloud players—Amazon, Microsoft, and Google—have all started pushing their “machine learning as a service” (MLaaS) solutions. Running the models in the cloud guarantees high scalability, but may fail to meet delay constraints. As an example, already deployed applications, such as recommendation systems, voice assistants, and ad-targeting, need to serve predictions from ML models in less than 20 ms [simsek16]. Future wireless services, such as connected and autonomous cars, industrial robotics, mobile gaming, augmented/virtual reality, have even stricter latency requirements, often below 10 ms and below 1 ms for the so-called tactile internet. It is then imperative to run these services closer to the user at the network edge. 5G deployment can provide computing and storage capabi