Скачать с ютуб ACACES 2024: Hardware/software co-optimization for machine learning at the edge, Lecture 1 в хорошем качестве

ACACES 2024: Hardware/software co-optimization for machine learning at the edge, Lecture 1

'Quantity has a quality of its own', says Giovanni Ansaloni (‪@epfl‬ ) – so how do you squeeze artificial intelligence (AI) into power-constrained edge devices? In his first lecture from ACACES 2024, Giovanni lays out the context for recent developments in AI, before explaining some of the potential of edge AI. Highlighting the tension between the massive data requirements of AI and the power and size limits of edge devices, he goes into some of the issues in implementing machine learning at the edge, including sparsity, event-based sampling and model optimization. This lecture focuses on 'what to compute'; subsequent lectures will cover 'how to compute'. 00:00 Intro 06:20 AI and edge 11:30 Bio-signal databases (SzCore) 15:05 Signal compression leveraging inputs sparsity and self similarity 27:20 Lightweight feature extraction and classification 52:30 EdgeML model optimization (few-shot learning, personalization, distillation, federated learning) ACACES 2024 was a joint summer school organized by HiPEAC (https://www.hipeac.net) and DISCOVER-US (https://discover-us.eu). Further information: ACACES 2024: https://www.hipeac.net/events/#/acaces/ Giovanni Ansaloni's EPFL webpage: https://people.epfl.ch/giovanni.ansal...