The Cloud-Edge AI Continuum

While deep ML models have been designed and operated mostly in the cloud, where there is enough computation power for their needs, we are starting to see a proliferation of ML models at the edge, on lower power devices. Increasing privacy awareness and latency are driving this trend of consuming data closer to its source.

Model compression and neural architecture search have been some of the most common approaches for providing ML models ready for the edge. But no matter how hard we try, producing tight models for the more resource constrained devices will not be possible without compromising on accuracy.

In this talk, Valentin will focus on hybrid inference through splitting the model computations between the edge and the cloud. We ensure user data privacy by running the early computations of the model with raw data only on the edge. We will explore solutions for model inference but also training between the edge and cloud.

Meet the speaker

Valentin Radu
Valentin Radu Applied scientist at Amazon

Valentin is a applied scientist at Amazon, in the Ring AI organisation, working on efficient computer vision models designed for edge computing in Ring security cameras. Previous to joining Ring, Valentin worked at Samsung, Intel and in the UK academia. His research evolves around efficient ML systems for inference and training, multimodal AI and context recognition. In his free time he enjoys mountain biking and hiking.