Stylianos I. Venieris’ Post

Happy to announce that in 2025's edition of Design, Automation and Test in Europe Conference (DATE), I'm joing Christos Kyrkou to co-chair a new topic area: we introduce topic "A6 - Applications of Artificial Intelligence Systems". Our core focus is novel design methods and optimizations across the stack for domain-specific AI applications, targeting any platforms in the computing spectrum (from microcontrollers to data centers). We are looking for full-length papers (6 pages + 1 for references only). For more details, checkout the conference website: https://lnkd.in/emgvEyJa Submission site: https://lnkd.in/e2KPxsdj

Earlier this week, I had the pleasure to give a keynote talk at #MobiUK 2024, discussing Consumer Edge AI 2.0, a novel computational paradigm that places EdgeAI-Hubs at the forefront in order to address the challenges of deploying contemporary and future AI use cases close to users. Our work has been accepted by IEEE Pervasive Computing Magazine. Congratulations and many thanks for the great work to Stefanos Laskaridis, Alexandros Kouris, Rui Li and Nicholas Lane. Extended preprint: https://lnkd.in/dSbT2tNH Paper: https://lnkd.in/d4k8yhw6

Yesterday, I had the chance to exchange insights with a great set of panelists at the AI Hardware & Edge AI Summit in London. Great discussion on the unique systems challenges of GenAI and the work we need to do to bring it closer to the users. Many thanks to Sally Ward-Foxton (EE Times | Electronic Engineering Times) for the excellent questions and to Jake Kochnowicz (Imagination Technologies), Phillip Burr (Lumai) and Roberto Mijat (Blaize) for the fruitful discussion. AI Hardware & Systems Samsung Electronics

Personalising AI models to the specific needs of a user and to each new environment -and doing so solely with on-device resources- is a key enabler towards richer, but also privacy-preseving, new experiences. While existing efforts have primarily focused on accommodating the resouce constraints of ubiquitous computing devices, such as microcontroller-based IoT modules, through more compute- and memory-efficient training methods, we are still challenged by the significant barrier of the scarce (labelled) data availability on devices in the wild, which puts a hard limit to the effectiveness of personalisation. In our paper, accepted by #ICML 2024, we present TinyTrain, an on-device training method that enables fast on-device personalisation on data-scarce, low-end platforms. This is enabled by: 1. A new sparse-update approach that fine-tunes only part of the model's parameters. Taking parameter-efficient fine-tuning (PEFT) methods a step further, we show that selecting the parameters to be fine-tuned in a dynamic and task-adaptive manner yields rapid and high-accuracy personalisation despite the low data availability. 2. A multi-objective parameter selection criterion that co-optimises accuracy, compute and memory footprint, specially designed for resouce-constrained platforms. Deployed using our in-house on-device training engine, TinyTrain delivers 9.5x faster and 3.5x more energy-efficient personalisation over status-quo fine-tuning and transfer-learning approaches, while achieving 2.23x smaller memory footprint than the state-of-the-art efficient on-device training methods. Congrats to Young D. Kwon for leading the work and to the rest of the great collaborators Rui Li, Jagmohan Chauhan, Cecilia Mascolo, Nicholas Lane. Preprint: https://lnkd.in/d3R2pkMf Codebase: https://lnkd.in/ePK-xi4n (soon to be open-sourced)

Our Generative AI (GenAI) group at Samsung AI in Cambridge, UK is growing and hiring #ML Engineers. As part of our Embedded AI division, the GenAI group works closely with the Speech group and our Distributed AI team, to conduct cutting-edge research and make it deployable across millions of Samsung devices. If interested, you can contact any of Sourav Bhattacharya, Abhinav Mehrotra or Alberto Gil Ramos. See below for further information and contact details.

Following up on our MultiTASC work (#ISCC 2023), we now present MultiTASC++, a next-generation scheduler for Multi-Device Cascades for DNN inference. MultiTASC++ introduces a number of techniques that enable the continuous adjustment of scheduling decisions, leading to dynamic adaptation to sudden changes, better scalability to more client devices, and sustained performance under various scenarios. The paper has been accepted at ITU's Journal on Future and Evolving Technologies (ITU J-FET). Congrats to Sokratis Nikolaidis for leading the work and to the ICBNet lab at National Technical University of Athens. Paper: https://lnkd.in/dEaC8uuU

I'm excited to announce that I'll speak at the AI Hardware & Edge AI Summit Europe, in a panel discussing Next-Generation Hardware & Systems Design for AI. After 6 successful years in Silicon Valley, the AI Hardware & Edge AI Summit will make its European debut in London, June 18-19, 2024. Focusing on making AI faster, more efficient, and more affordable, with sessions covering the latest developments in: - Server/Cloud AI Infrastructure - Industrial Edge AI Infrastructure - Hardware & Systems - Software Infrastructure & ML Applications This promises to be an outstanding event that brings together the entire ecosystem. If you'd interested in joining me, you can register today and quote SPEAKER10 to save 10%, following this link: https://lnkd.in/dsnnPn7p Details of the agenda: https://lnkd.in/d4RWvMAF #AIHardware #AIHardwareSummit #EdgeAI #EdgeAISummit #Hardware #AIHardwareConference #AI #AIConference

Multi-DNN workloads are rapidly becoming an emergent workload across mobile and server inference systems. In our paper accepted by #MICRO 2023, we present Dysta, a new scheduling/hardware co-design approach that pushes the attainable performance of NPUs under ML model multi-tenancy. In this work, we introduce two key techniques: 1) at the scheduling level, a new scheduling algorithm that exploits both static and dynamic sparsity of each infererence request; we showcase how by considering the unique sparsity features of incoming inference requests, such as the sparsification pattern, more precise and performat scheduling decisions can be made. 2) a hardware-based dynamic scheduler that monitors and adapts the execution with low resource overhead, enabling efficient preemption, minimal interruption of the NPU, and improved satisfaction of the latency SLOs. In the paper, we demonstrate Dysta's robustness across various levels of request traffic, and show how it achieves significant performance gains (improved tail latency and hardware utilisation) across both CNN- and Transformer-based multi-DNN workloads. We are further pleased to announce that we have open-sourced the associated code (see the link below). Many thanks to Hongxiang Fan for co-leading this work and to the rest of the authors Alexandros Kouris and Nicholas Lane. Preprint: https://lnkd.in/evRP5yXH Codebase: https://lnkd.in/euUKzh3x