Speakers
Transforming Industries with Edge AI — Session 2, 11:15–11:30
Edge AI is transforming industries by enabling distributed computing and intelligent applications directly at the point where data is generated and actions are performed. As AI models become increasingly complex and computationally demanding, the need for higher performance at the edge continues to grow. At the same time, edge systems must preserve their core characteristics—low power consumption, minimal latency, and efficient real-time processing. Balancing these requirements is critical to unlocking the full potential of next-generation Edge AI solutions. In this session, you will see how Axelera AI is supporting the balance with some real-life use cases.
Edge AI for Anomaly Detection and Self-Correction in Smart Lighting Systems: Opportunities and Challenges — Session 2, 11:30–11:45
Smart lighting systems depend on multiple heterogeneous sensors and wireless protocols to autonomously control light output. The application logic for light control is mainly through embedded state machines and algorithms that run on resource-constrained microcontrollers. In real-world deployments, these systems face degraded behavior caused by faulty sensor readings, unhandled corner-case scenarios in multi-source occupancy fusion, and wireless packet loss due to RF interference and network congestion. This talk explores the opportunities and challenges of leveraging Edge AI techniques to detect and correct such anomalies directly on the IoT node, without cloud connectivity. We discuss lightweight on-device strategies for identifying deviations from expected behavior, arbitrating conflicting sensor inputs, and enabling graceful recovery under degraded conditions. Such models are also expected to fit within the tight memory, compute, and real-time constraints of deployed IoT luminaires.
Live acoustic monitoring at the edge with microphone arrays — Session 2, 11:45–12:00
There is something elegant about acoustic AI at the edge: the same deployment pipeline can connect domains as different as stadiums, industrial plants (with robots), and airports. This talk presents how microphone arrays and NVIDIA Jetson platforms can be used to perform live acoustic monitoring with low latency and actionable outputs. Using examples from these environments, we show how sensing, signal processing, and AI inference are combined in a practical edge deployment pipeline. The talk highlights how these seemingly unrelated domains share common requirements in latency, robustness, and scalability, and what this means for the design of embedded acoustic AI systems.
Hardware-Aware Low-Power Object Detection on STM32U5 — Session 3, 13:00–13:15
This talk presents an optimized edge AI system for object detection in digital agriculture, based on the YOLOv8n object detector deployed on the STM32U575ZI microcontroller, with a specific emphasis on low power consumption. Several compression techniques are applied to the detection model, including structured pruning, integer quantization and input scaling in order to meet strict hardware constraints. The model is trained and evaluated on the CropAndWeed and Lincoln Beet datasets, achieving a balanced trade-off between detection accuracy and efficiency.
Model and AI Inference Optimization for Resource-Constrained Edge AI Devices — Session 3, 13:15–13:30.
Edge AI is becoming essential for delivering intelligent, responsive, and privacy-preserving applications directly on devices, without relying on continuous cloud connectivity. However, running increasingly diverse and complex AI models at the edge requires state-of-the-art model and inference optimization methods to meet strict constraints on latency, power, memory, and cost. In this talk, we discuss how NXP applies and develops state-of-the-art model and AI inference optimization techniques for efficient execution on NXP Edge AI NPUs. We will cover techniques ranging from model quantization and model and token pruning to speculative decoding, showing how these techniques help enable scalable deployment of a wide range of AI models on embedded platforms.
Compute-in-Memory (CIM): A Key Enabler for Edge AI — Session 3, 13:30–14:00
Compute-in-Memory (CIM) has emerged as a promising paradigm to overcome the energy and bandwidth bottlenecks of conventional von Neumann architectures, making it particularly attractive for Edge AI. This talk introduces the fundamentals of CIM and reviews the state of the art in performance, efficiency, and scalability. Next we discuss key CIM variants and their application to both artificial neural networks (ANNs) and spiking neural networks (SNNs). Finally, we outline future trends and architectural directions for extending CIM toward more demanding workloads, including large language models (LLMs), in resource-constrained edge environments.
Emerging Computing Technologies for Neural Interfaces — Session 4, 14:15–14:45
Emerging Computing Technologies for Neural Interfaces: This talk explores the development of brain-inspired brain–computer interfaces (BCIs), covering both circuits and systems for neural interfacing and computing architectures implemented in digital, analog, and emerging technologies such as spintronics, memristors, and CMOS. It highlights applications in brain monitoring and treatment of neurological disorders like Parkinson’s disease and epilepsy. The central focus is on demonstrating the feasibility and potential of neuromorphic computing systems to efficiently process and interface with neural bio signals, enabling next-generation, low-power, and intelligent BCI solutions.
Doing More with Less: Smart Model Compression for Efficient Edge AI — Session 4, 14:45–15:00
The rapid expansion of Edge AI demands intelligent systems that operate within energy, memory, and computational budgets. Although specialized hardware accelerators provide a foundation for efficiency, hardware optimization alone can no longer keep pace with the exploding complexity of modern neural networks. This presentation shifts the focus to the algorithmic frontier, introducing a novel, lightweight framework for joint pruning and quantization. Powered by uncertainty-driven sensitivity scores, this approach eliminates the need for isolated, sub-optimal optimization steps or brute-force search. Crucially, the tool itself features minimal execution time and memory overhead, making the compression process highly efficient. Through experimental results on various datasets, we will demonstrate how this framework drastically reduces model footprints without sacrificing accuracy, unlocking the next generation of scalable and sustainable Edge AI.
Real-Time FMCW Radar Estimation: Chirp-by-Chirp Signal Processing on Neuromorphic Hardware — Session 4, 15:00–15:15
Conventional FMCW radar processing is limited by frame-based FFTs that require large data buffers and introduce significant latency. This talk presents an event-driven alternative that applies neuromorphic principles to achieve efficient perception at the edge. By using a two-stage Spiking Neural Network (SNN), we process radar signals chirp-by-chirp: Integrate-and-Fire neurons perform initial range estimation, while Resonate-and-Fire neurons extract velocity through frequency-tuned oscillations. This architecture enables high-velocity targets to be detected with fewer chirps, significantly reducing detection latency compared to standard pipelines. We demonstrate how these spike-based strategies, implemented on specialized neuromorphic hardware, overcome the memory bottleneck of traditional 2D FFTs, providing a low-power, low-latency solution for real-time edge AI
Synaptic delays for embedded AI models — Session 4, 15:15–15:30
Delay-parameterization of ML models has been a consistently re-appearing theme in the literature of neural models. The recent re-emergence has brought advances in training methods, as well as consistent evidence of performance improvement in deep models. However, as delay parameters scale with the number of synapses, one naturally may ask, "What about hardware efficiency?" In this presentation, we present empirical insights on this question, based on the pursuit of minimum-footprint, delay-parameterized models for the MLPerf-tiny KWS benchmark.
Interactive Edge Intelligence for Next-Generation Human–Machine Systems — Session 5, 15:45–16:00
The next generation of interactive edge intelligence systems, such as smart glasses and XR headsets, robots, requires seamless integration of perception, interaction, and intelligent processing. This talk presents a unified framework that bridges human–machine interaction with efficient on-device intelligence. It highlights cost-effective interface designs, including event-based eye tracking, hand tracking, and voice control, alongside energy-efficient edge intelligence solutions such as neuromorphic AI hardware and compact large language models for real-time operation under strict energy and latency constraints. A case study further illustrates how eye gaze signals can guide visual–language models (VLMs), reducing computational workload while enabling adaptive and context-aware interaction.
Embedded AI for Computer Vision and Biometrics Applications — Session 5, 16:00–16:15
Computer Vision and Biometrics application, like e.g. face recognition, are typically resource hungry. On the one hand, deep learning and AI developments have drastically improved the performance of image analysis and biometric recognition. On the other hand, this has resulted in a significant increase of the required resources. However, there is also a need for compact and stand-alone systems for image analysis. In the Computer Vision and Biometrics LAB at the University of Twente, we investigate and develop such systems. We develop stand-alone small factor systems for e.g. finger vein and iris recognition that use modest computing platforms like the Raspberry PI. Another application is the use of AI methods for automotive applications like Self Driving Vehicles. In addition, we investigate methods to shrink and adapt large AI models for Computer Vision tasks and Biometric Recognition to fit on modest platforms, like FPGAs. This is a multi-objective optimisation problem that aims to preserve accuracy and throughput as much as possible while reducing memory and computing resources. One of the used techniques is called Network Architecture Search, where the best architecture for a specific platform is found.
A sensor you can swallow: highly miniaturized ingestible technology for real-time, patient-friendly insights into gut health — Session 5, 16:15–16:30
The human gastrointestinal tract is crucial for our overall well-being, yet it remains a black box to a certain extent. Current gold-standard diagnostic tools like endoscopy are invasive and require uncomfortable bowel preparation, while fecal tests offer only partial insights into gastrointestinal (GI) function. Ingestible technology has the potential to unravel this black box by providing a detailed, non-invasive view of the entire GI tract. At the OnePlanet Research Center / imec, two such technologies are being developed: the Gastrointestinal Smart Module (GISMO), a sensing pill, which measures pH, temperature, redox potential, and regional transit times; and a sensor-guided sampling pill that collects targeted intestinal content for offline analysis. These devices were thoroughly validated in pre-clinical in vitro and in vivo models. The sensor pill was successfully tested in a first-in-human trial involving 15 healthy participants and is currently deployed in patient trials. Together, these innovations represent a major step forward in GI diagnostics and monitoring, offering real-time, patient-friendly insights into gut health.
Neuronflow: Low Power Low Latency AI for Responsive See-through AR— Session 4, 16:30–16:45
This talk presents an overview of the machine learning training and neural network design techniques used to exploit the sparse execution capabilities of NeuronFlow, a brain-inspired, event-driven neural processing unit (NPU) designed for low-power, low-latency operation in see-through augmented reality (AR) glasses.