Invited Speakers

Merging Insights from Artificial and Biological Neural Networks for Neuromorphic Edge Intelligence

Abstract: The development of efficient bio-inspired algorithms and hardware is currently missing a clear framework. Should we start from the brain computational primitives and figure out how to apply them to real-world problems (bottom-up approach), or should we build on working AI solutions and better align them with key neuroscience insight (top-down approach)?

In this talk, we will see why biological plausibility and hardware efficiency are often two sides of the same coin, and how neuroscience- and AI-driven insights can cross-feed each other toward neuromorphic edge intelligence, which I will highlight with our latest chip ReckOn, a 0.45-mm² spiking recurrent neural network processor enabling end-to-end task-agnostic online learning over second-long timescales. Benchmarked on vision, audition and navigation tasks within a learning power budget of 50µW at 0.5V, ReckOn illustrates how a neuromorphic approach can bring unique advantages for edge devices that continuously adapt to their environment.

Danilo Pau

(STMicroelectronics)

Technologies for In Sensor Tiny Machine Learning Computing

Abstract: Deep learning frameworks have been focused on floating point computation. Since images are typically represented with 8bits per pixel, most of these tools support post-training quantization procedures with marginal accuracy drops. However, heterogeneous sensors do not produce necessarily that bit representation since they shall provide adequate signal to noise ratio with proper bit-depth per sample.

Starting from this consideration, new deep learning frameworks have been introduced to support fractional numbers representation to match sensor data precision and associated pre-processing. Furthermore, addressing low memory footprint, efficient computing, high accuracy, and low power consumption is a challenge when designing neural networks with fractional arithmetic for tiny, embedded devices. These Deeply Quantized Neural Networks (DQNNs) offer the most interesting approach with non-trivial design and training processes. They use hybrid precision variants developed with research deep learning tool like QKeras, and can achieve interesting accuracies compared to more traditional design approaches with lower memory footprint.

Applying these neural networks, so much hyper-parametrized, on use cases spanning from anomaly detectors to classifiers design shall also achieve high accuracy and be deployable on micro-controllers and sensors. Since standard off-the-shelf microcontrollers do not feature dedicated instructions for DQNN execution, code generation tools (such as X-CUBE-AI) have been conceived to output optimized code that can run efficiently on these execution targets.

In this presentation, all those aspects will be discussed with reference to the latest efforts of STMicroelectronics, including X-CUBE-AI for efficient deployment on micro-controllers and sensors.

Federico Corradi

(TU Eindhoven)

Asynchronous Low-power Spiking Neural Network Processing Systems for Extreme-edge Applications

Abstract: Without a doubt, we are still many orders of magnitude away from reaching the incredible efficiency, speed, and intelligence found in natural brains with our modern computing architectures. Moreover, our traditional computing systems are now hitting hard limits, such as the memory wall and the end of Dennard scaling (i.e., the performance per watt increase has slowed significantly). For these reasons, many scientists worldwide are researching new computing architectures inspired by the brain. In this regard, neuromorphic computing is among the most promising approaches for achieving energy-efficient hardware systems for real-time signal processing. Thus, it has the potential to enable several edge artificial intelligence tasks. In this approach, brain computation is mimicked at the circuit level, employing event-driven and massively parallel spiking neural networks directly implemented at the hardware level.

During this talk, I will present the computing paradigm of spiking neural networks and illustrate novel training algorithms and digital computing architectures based on ultra-low-power and massively parallel implementations of neurons and synapse circuits. Finally, I will showcase a few prototype devices that can meet extreme-edge applications' strict energy and cost reduction constraints on the Internet of Things and for biomedical signal processing applications.

Jie Yang

(TU Delft)

Human-Centered Computing for Robust Embedded AI

Abstract: Embedded Artificial Intelligence is used more and more in society, from healthcare to transport and smart buildings. Along with the rapid increase of Embedded AI adoption comes increased concerns about the inherent robustness issues of such technologies and the social, and ethical implications. To create Embedded AI systems that can properly serve humans, it is crucial to put humans at the center of the process such that the outcome system behaves in a way that fits the values and needs of people in the contexts of use. This poses new challenges to technological development: how to build Embedded AI systems that can be understood by humans and that can align their behaviour with human values? Tackling these challenges requires new ways of looking at Embedded AI systems, e.g., machine learning models as knowledge bases and as autonomous agents that people can query, interact with, and influence. In this talk, I will present our recent work on understanding and improving the robustness of embedded AI systems in both centrealised and federated learning scenarios.

Manuele Rusci

(KU Leuven)

From Deep Inference to On-Device Learning with Multi-Core RISC-V MicroControllers at the Extreme-Edge

Abstract: This talk focuses on the methodologies for designing low-power smart sensors for the extreme-edge of the network. Because of the severe energy constraints, MicroController (MCU) devices are typically used for peripheral control and data analytics. By means of practical examples, e.g., smart cameras with object detection capabilities, autonomous nanodrones or tiny hearables featuring speech enhancement, this talk discusses tools and software architectures designed to efficiently bring Deep Learning based inference tasks onto state-of-the-art multi-core MCUs with extended RISC-V cores, along with (mixed) low-precision quantization methods for Deep Learning models. Lastly, we review the initial support for on-device learning with the perspective of turning these static low-power smart sensors into a distributed network of MCU-based sensors able to continually adapt or learn from the surrounding environment.

Stefanos Laskaridis

(Samsung AI Center)

Capturing and Tackling System Heterogeneity in the Wild: Tales from Deploying DNNs across Various Devices

Abstract: Smart devices are omnipresent nowadays, and a valuable assistant to human activity. They become more and more capable in their attempt to become smarter. However, such devices come in various forms and flavours and co-exist with devices of different tiers, budgets and generations. The result is that a computationally heterogeneous ecosystem is shaped, responsible for handling a variety of AI tasks in the wild, spanning across inference and training workloads. How one can elastically deploy models in such an environment, collaborative or not, and dynamically tune the workload to the device at hand is an open challenge.

During this talk, I will present our work in the field of measuring and tackling heterogeneity. Specifically, I will talk about our attempts to measure and understand this computational discrepancy for state-of-the-art and popularly deployed models and invent ways to efficiently run DNNs on devices of various capabilities for inference and federated training across a variety of tasks and modalities. Last, noticing the trajectory of the AI landscape, with the advent of LLMs, the metaverse and diffusion models, I will talk about how we set the foundations for a new paradigm of organising workloads at the edge.