ENGAGE is a project within the framework of the strategic research and innovation agenda of DFKI and Inria in the field of artificial intelligence. Advances in machine learning algorithms and hardware have allowed Deep Neural Networks (DNN) to become a pervasive tool across a large range of industrial and scientific domains. However, the availability of training data in sufficient quantity and quality remains a central issue. In the future, data will increasingly be generated on-the-fly using parametric models and simulations (in-silico). This is particularly useful in situations where obtaining data by other means is expensive, raises ethical concerns, or where a phenomenon has been predicted in theory, but not yet observed. One key application of this approach is to validate and certify AI systems through targeted testing with synthetically generated data. The approach solves most data-related issues with the application of DNN in real-world industrial or scientific scenarios but raises questions on all levels of existing DNN infrastructure. The project addresses the question how the adaptive sampling of the parameter spaces of complex models will allow for better choices on what data to generate and how a model decay can be recognized. In order to distribute trained models ENGAGE address the question how virtualization and scheduling need to be adapted to facilitate the resulting mixed workloads consisting of training and simulation tasks and running on federated HPC/cloud/edge infrastructures. On the resource management level, ENGAGE contributes novel strategies to optimize memory management, the dynamic choice of the parallelization methods and secure cross hardware (CPU, GPU, FPGA, etc.) deployment strategies using the AnyDSL compiler environment. Targeted fields of application are Industry 4.0, Smart Living, Smart Energy networks and autonomous driving.
Institut national de recherche en sciences et technologies du numérique (INRIA)