These are the abstracts of the different talks:

Bernabé Linares

Event-driven convolution based processing

Abstract: We review some of the event-driven hardware developments in which our lab has been involved, covering from sensitive-DVS to event-driven convolutions on dedicated ASICs, FPGAs, and the SpiNNaker platform, with applications in object recognition or stereo vision. We show how to train event-driven convnets to minimize the number of required spikes, reducing energy consumption for the same recognition tasks. Additionally, we present some results on a type of spike-timing-dependent-plasticity, which uses only binary weights combined with stochasticity, and which results in hardware that requires less hardware and energy resources for the same accuracy.

Michael Hopkins

The SpiNNaker neuromorphic architecture and its application to a novel learning algorithm

Abstract: TBD.

Daniele Ielmini

Emerging resistive switching devices for neuromorphic sensing and learning

Abstract: Emerging resistive/memristive memory devices can enable brain-inspired computing primitives, thanks to their tunable volatile/nonvolatile characteristics, their excellent scaling in both 2D and 3D, their ability to run in-memory compute algorithms, and their unique physical properties that can mimic the individual building blocks in the brain, such as neurons, synapses and dendrites. Exploring alternative materials/devices may disclose novel physical phenomena and different scales of voltage, current and time, thus providing new opportunities for neuromorphic engineering technologies. This talk will present the status and challenges about resistive switching devices for brain-inspired spike-based sensing and computing. A broad scope of emerging devices will be illustrated, including volatile switching devices, memtransistors devices based on 2D semiconductors such as MoS2, memristive nanowire structures and electro-chemical transistors. The prospects for hybrid memristive-CMOS circuits for neuromorphic computing will be discussed in terms of scaling and energy efficiency.

Alon Ascoli

Locally-Active Memristors Allow to Reproduce High-Order Biological Phenomena through Low-Order Bio-Inspired Dynamic Arrays

Abstract: The recent introduction of disruptive miniaturized volatile memristor devices, blessed with the capability to act as small-signal energy sources under suitable polarization, similarly as the sodium and potassium ion channels in the axon membrane, enables the design of bio-inspired circuits capable to reproduce the rich nonlinear dynamics of biological systems. Recurring to the vast body of knowledge in circuit and system theory, the subsequent analysis of these memristor-based networks allows to gain a deep insight into the mechanisms underlying the emergence of complex dynamical phenomena in the original biological systems. This seminar first reveals the capability of a volatile threshold switch, manufactured at NaMLab, to amplify infinitesimal fluctuations in energy. The local activity of the miniaturized device is then leveraged to design bio-inspired cellular neural networks, which require a low number of degrees of freedom to reproduce diffusion-driven instabilities in higher-order biological reaction-diffusion systems. An in-depth theoretic investigation of the local and global dynamics of the proposed circuits sheds light into the origin for complexity in the biological systems, resolving a dilemma, which troubled the minds of luminaries of the calibre of Turing, Prigogine, and Smale for decades.

Catherine Schuman

Algorithms for Neuromorphic Computing and Neural Hardware

Abstract: Neuromorphic computing and neural hardware are popular technologies for the future of computing. Much of the focus in neuromorphic computing and neural hardware research and development has focused on new architectures, devices, and materials, rather than in the software, algorithms, and applications of these systems. In this talk, I will overview these fields from the algorithm perspective. I will specifically highlight how machine learning algorithms have to be adapted for these new hardware systems to use them most effectively, and I will discuss how non-machine learning algorithms can be mapped onto some of the hardware systems as well.

Luis Hernández

Hands-on Deep Learning with Keras and Tensorflow

Abstract: In this workshop you will learn to develop Deep Learning models using Keras on top of Tensorflow. Keras is one of the leading high-level neural networks APIs. Through an image classification case study, you will be able to define, train, evaluate, understand and visualize different basic Deep Learning architectures. From these simple examples, we will discuss more advanced topics, such as attention mechanism, transformers, transfer learning or self-supervised learning.  In the workshop we will use Python Notebooks running under the Google Colaboratory free environment.

Michele Rossi

Energy Efficient Scheduling at the Network Edge: Current Solutions and Future Research Pathways

Abstract: There is a current trend towards moving the computation processes generated by end devices toward the mobile network edge. This offers advantages such as reduced latencies for the end users, but also relieves the mobile and Internet networks from the burden of carrying traffic to be processed at network cloud facilities. The latter aspect is particularly appealing for two reasons: i) a reduced energy consumption for the cloud computing servers, including their energy hungry cooling systems, ii) the possibility of exploiting renewables at the network edge, so to potentially power computing processes out of green energy. Taking this setup into account, in this talk we present scheduling solutions to balance computing resources at the edge servers, on the fly, and according to several factors such as end-user computation deadlines, renewable energy availability at the edge servers and user mobility. Recently proposed mobility predictors will be presented, detailing their use within centralized and distributed scheduling frameworks, summarizing their performance and identifying future research avenues.

Marco Miozzo

Taming the Environmental Impact of AI with Collaborative and Distributed Learning at the Edge

Abstract: The amount of computations required to train Deep Learning (DL) models has increased 300,000 times during the last 6 years and the estimation for the near future maintains this trend. AI, together with mobile networks and services, is one of the main drivers of the upsurge of the energy footprint of Information and Communication Technology (ICT). In order to address a net zero target, as suggested by the recent EU Climate Pact, researchers and developers of AI would need to change their paradigm from accuracy to efficiency. To do so, collaborative and distributed learning represents key enablers for reducing the ICT energy consumption. In this talk, state-of-the-art solutions will be outlined, concentrating on centralized and peer-to-peer Federated Learning, Knowledge Transfer Learning and Continual Learning.

David Atienza

Edge AI Biomedical Systems for a Privacy-Preserving IoT Era

Abstract: The Internet of Things (IoT) has been hailed as the next frontier of innovation where our everyday objects are connected in ways that improve our lives and transform industries, in particular healthcare. In this talk, Prof. Atienza will first discuss the challenges of ultra-low power (ULP) Multi-Processor System-on-Chip (MPSoC) design and communication in edge Artificial Intelligence (AI) nodes for the design of biomedical devices and wearables in the IoT context. Then, the opportunities for edge AI architectures to conceive the next generation of federated learning systems in healthcare will be highlighted as a scalable way to deliver the IoT concept in a privacy-preserving way. This new trend of edge AI-based MPSoC architectures will need to combine new ULP heterogeneous embedded systems, including reconfigurable neural network accelerators, as well as enabling energy-scalable software layers. The final goal is to have edge AI systems that can gracefully adapt the energy consumption and precision of the IoT application outputs according to the quality requirements of our surrounding world. Moreover, they need to be able to personalize their AI algorithms by enabling training on the edge, as living organisms do to operate efficiently in the real world.

Josep Maria Margarit

Pushing the Analytical Limits through Fully-Integrated Electrochemical Sensing and Neuro-Inspired Perception

Abstract: The technological bloom of integrated microanalytical systems is poised to enable ubiquitous (bio)chemical fluid assessment, and to have a revolutionary impact on the prevention of key health and sustainability threats of our time. In these systems, the use of electrochemical sensor arrays stands out due to their capability to generate multivariate data from liquid samples, enlarging the number of chemical properties that can be determined simultaneously. To manufacture the arrays, microsensors fabricated in semiconductor technologies offer advantages such as miniaturization, robustness, mass fabrication, and ease of integration with electronic circuits for embedded data processing/storage/transmission, making them particularly suitable for advanced on-site monitoring. Even though the resulting smart microsensing devices are sensitive enough for a wide range of applications, their practical usage is still hampered by drift and cross-sensitivities that impair performance in real settings. This drawback adds to current limitations on modeling power to extract chemical information in situ and in real time. In this talk, Dr. Margarit will discuss electrochemical methods, sensors, and challenges for continuous on-chip analysis of aqueous media. He will also highlight neuro-inspired processing techniques to leverage chemical perception and to boost accuracy, autonomy, and portability in analytical devices.