P. Coscia, A. Genovese, V. Piuri, F. Rundo, F. Scotti, Tree-based optimization for image-to-image translation with imbalanced datasets on the edge, Proc. of the 2023 IEEE Int. Conf. on Computational Intelligence and Virtual Environments for Measurement Systems and Applications (CIVEMSA 2023), pp. 1-6, June 2023, ISSN 979-8-3503-3636-8.
Industrial informatics
Edge AI Technologies for Optimised Performance Embedded Processing
Edge AI
The EdgeAI project—Edge AI Technologies for Optimised Performance Embedded Processing—addresses the growing need to execute advanced artificial-intelligence functions directly on embedded and edge devices. Moving computation closer to sensors and industrial processes can reduce communication latency, improve responsiveness, limit the transmission of sensitive data, and support operation in environments where cloud connectivity is unavailable or undesirable. At the same time, edge platforms impose strict constraints on computing power, memory, energy consumption, model complexity, and execution time.
The project develops end-to-end hardware and software solutions covering the complete AI design stack and middleware. Its objectives include scalable and energy-efficient AI methods, support for heterogeneous operating systems and hardware platforms, advanced multicore systems-on-chip and systems-on-module, embedded hybrid architectures, industrial IoT connectivity, modular hardware/software co-design, and reconfigurable platforms that improve the reusability, updatability, and operational lifetime of AI systems.
EdgeAI operates across the micro-, deep-, and meta-edge, selecting the most appropriate execution level according to latency, reliability, energy efficiency, processing requirements, and application performance. The developed technologies are validated through industrial demonstrators organised into five value chains: digital industry, energy, agri-food and beverage, mobility, and digital society. These activities target applications such as advanced sensing, automated defect classification, virtual metrology, intelligent production control, distributed monitoring, perception for autonomous systems, and human-centred digital services.
A major research contribution concerns the optimisation of deep-learning models for devices with limited resources. One investigated approach combines several pretrained image-to-image translation networks through a binary-tree procedure that averages selected model weights and subsequently fine-tunes the resulting networks. This strategy is designed for scenarios involving strongly imbalanced datasets, limited availability of samples, privacy restrictions, and insufficient computational resources. Experiments show that the procedure can improve the perceptual quality and consistency of generated images in several highly imbalanced translation tasks.
The same research investigates quantisation-aware training for converting models from 32-bit floating-point representations to 8-bit integer execution. The results indicate that fine-tuning after quantisation is important for limiting the expected performance degradation, thereby supporting more efficient deployment on embedded and edge hardware. The work also identifies remaining challenges, including the computational cost of repeated fine-tuning and reduced benefits for image domains containing particularly complex patterns.
A second central theme is trustworthy and explainable Edge AI. Edge models must not only provide accurate predictions but also make their decisions understandable, particularly in safety-critical or high-risk applications. The project studies interpretability and explainability techniques, visual explanations, model-specific and model-agnostic approaches, knowledge extraction, and benchmarking methods. It also analyses the distinction between explaining individual predictions and understanding the internal mechanisms of an AI model.
The research highlights that explainability at the edge introduces additional trade-offs among transparency, predictive performance, model complexity, execution speed, memory requirements, and energy consumption. It also identifies the lack of standardised definitions, mature benchmarking procedures, and generally accepted evaluation measures as important open problems. The project therefore promotes explainability-by-design, rigorous model evaluation, data-quality validation, and hybrid approaches that balance technical performance with transparency and user trust.
Overall, EdgeAI contributes technologies and methodologies for making embedded AI more efficient, secure, adaptive, explainable, and deployable in real industrial environments. Its results support the transition from cloud-centred intelligence to distributed AI systems capable of performing real-time sensing, analysis, and decision-making directly where data are generated.
The project develops end-to-end hardware and software solutions covering the complete AI design stack and middleware. Its objectives include scalable and energy-efficient AI methods, support for heterogeneous operating systems and hardware platforms, advanced multicore systems-on-chip and systems-on-module, embedded hybrid architectures, industrial IoT connectivity, modular hardware/software co-design, and reconfigurable platforms that improve the reusability, updatability, and operational lifetime of AI systems.
EdgeAI operates across the micro-, deep-, and meta-edge, selecting the most appropriate execution level according to latency, reliability, energy efficiency, processing requirements, and application performance. The developed technologies are validated through industrial demonstrators organised into five value chains: digital industry, energy, agri-food and beverage, mobility, and digital society. These activities target applications such as advanced sensing, automated defect classification, virtual metrology, intelligent production control, distributed monitoring, perception for autonomous systems, and human-centred digital services.
A major research contribution concerns the optimisation of deep-learning models for devices with limited resources. One investigated approach combines several pretrained image-to-image translation networks through a binary-tree procedure that averages selected model weights and subsequently fine-tunes the resulting networks. This strategy is designed for scenarios involving strongly imbalanced datasets, limited availability of samples, privacy restrictions, and insufficient computational resources. Experiments show that the procedure can improve the perceptual quality and consistency of generated images in several highly imbalanced translation tasks.
The same research investigates quantisation-aware training for converting models from 32-bit floating-point representations to 8-bit integer execution. The results indicate that fine-tuning after quantisation is important for limiting the expected performance degradation, thereby supporting more efficient deployment on embedded and edge hardware. The work also identifies remaining challenges, including the computational cost of repeated fine-tuning and reduced benefits for image domains containing particularly complex patterns.
A second central theme is trustworthy and explainable Edge AI. Edge models must not only provide accurate predictions but also make their decisions understandable, particularly in safety-critical or high-risk applications. The project studies interpretability and explainability techniques, visual explanations, model-specific and model-agnostic approaches, knowledge extraction, and benchmarking methods. It also analyses the distinction between explaining individual predictions and understanding the internal mechanisms of an AI model.
The research highlights that explainability at the edge introduces additional trade-offs among transparency, predictive performance, model complexity, execution speed, memory requirements, and energy consumption. It also identifies the lack of standardised definitions, mature benchmarking procedures, and generally accepted evaluation measures as important open problems. The project therefore promotes explainability-by-design, rigorous model evaluation, data-quality validation, and hybrid approaches that balance technical performance with transparency and user trust.
Overall, EdgeAI contributes technologies and methodologies for making embedded AI more efficient, secure, adaptive, explainable, and deployable in real industrial environments. Its results support the transition from cloud-centred intelligence to distributed AI systems capable of performing real-time sensing, analysis, and decision-making directly where data are generated.