Many industrial actors have today not one, but two distinct ML departments - one mostly dedicated to R&D into advanced ML, and the other to the actual embedded implementation of ML algorithms. The embedded ML department usually considers rather simple (e.g. feedforward) networks in inference mode and focuses on providing real-time and numerical precision guarantees using compilers and run-times with limited control over optimization and resource allocation. But advanced ML algorithms require complex control involving back-propagation training (in on-device training and reinforcement learning contexts), or more generally stateful (e.g. recurrent) and conditional (e.g. gated mixture of experts) behaviors. Such advanced control is not readily covered by existing embedded back-ends, and is also often hidden in layers of Python code. We propose an approach to address and 

Transforming a trained model into an executable implementation must preserve the model’s intended semantics. Focusing on models specified using the ONNX standard, this presentation examines the requirements for precise and unambiguous syntax and semantics when models are interpreted or translated into lower-level representations. It highlights the limitations of ONNX in certification-oriented contexts and discusses the need for a safety-related ONNX profile. The presentation also introduces the objectives, methodology, and initial results of the SONNX Working Group, which aims to align the use of ONNX with industrial and safety-critical requirements without diverging from the standard.

Modern deep learning systems rely heavily on numerical approximations, including reduced precision, quantization, non-deterministic execution orderings, and parallel computation. These choices do not merely introduce “negligible noise”; they can fundamentally alter optimization dynamics, learned representations, training behavior, stability, and, in some cases, the functional behavior of neural networks. This presentation sheds light on the tensions between determinism, reproducibility, and predictability, and questions the actual role of numerical precision in the design, analysis, and certification of modern machine learning models.

The deployment of AI at the edge is often driven by performance constraints, with safety  considerations addressed only at later stages. Aidge challenges this paradigm by providing an open-source framework in which confidence, and traceability are first-class design objectives alongside performance optimization. 
It offers a transparent and auditable toolchain for deploying inference on edge platforms, relying on explicit intermediate representations and controlled graph transformations to ensure traceability from training models to implementation artifacts. Through the integration of the ACETONE approach, enabling traceability and worst-case execution time analysis, Aidge is well suited to aeronautical certification standards such as DO-178C and the forthcoming ML-specific standard ARP6983. In addition, Aidge supports inference testing under hardware fault conditions and pioneers the adoption of the Safety ONNX standard, contributing to robustness assessment, standardized model exchange, and strengthened assurance arguments for safety-critical AI systems.