tinyML Talks: tinyML: Designing Efficient Neural Architectures and Scaling Strategies for Edge…

tinyML: Designing Efficient Neural Architectures and Scaling Strategies for Edge Computing
Francesco Paissan
Junior Researcher
Energy Efficient Embedded Digital Architectures (E3DA) Unit
Fondazione Bruno Kessler (FBK)

In the Internet of Things (IoT) era, where we see many interconnected and heterogeneous mobile and fixed smart devices, distributing intelligence from the cloud to the edge has become critical for the sustainability of the infrastructures and comes with additional benefits (e.g. power efficiency, low-latency inference, privacy-by-design) with respect to centralized cloud computing. This paradigm brings new challenges in deep learning, such as low memory availability and limited energy budget. In this seminar, we will discuss some of our recent efforts to tackle the challenges of tinyML with novel neural architectures, training paradigms, and scaling strategies. In particular, we will focus on efficient multimedia analytics pipelines that achieve state-of-the-art results with a fraction of the computational budget of competitive approaches. Among the practical applications of these novel methodologies, we will discuss their performance for object detection, tracking, and zero-shot audio classification.

source