Speaker
Description
Title: Training ML algorithms on resource-constrained devices - a memory/storage perspective
Summray:
Deploying ML/AI algorithms on the edge is necessary for applications (e.g., security and surveillance,
industrial IoT, autonomous vehicles, healthcare use cases, ...) requiring low latency, data privacy or reduced costs. However, most edge devices are not equipped with powerful memory systems to perform such memory and processing intensive applications. The objective of this presentation is to show some optimization venues to unlock the memory/storage bottleneck of some ML/AI algorithms mainly from a learning perspective to deply them on low-resource devices. The optimizations presented in this talk could be also applied to whatever resource constrained device used for training, be it cheap virtual machines on cloud infrastructures, common personal computers or resource contrained micro datacenters.
Deploying ML/AI algorithms at the edge is essential for applications such as security and surveillance, industrial IoT, autonomous vehicles, and healthcare, which require low latency, data privacy, or reduced costs. However, most edge devices lack powerful memory systems capable of handling the memory- and computation-intensive nature of such applications.
The objective of this presentation is to highlight some optimization strategies that help overcome the memory and storage bottlenecks of ML/AI algorithms—mainly from a training perspective—to enable their deployment on low-resource devices. These optimizations can also be applied to any resource-constrained environment used for training, including low-cost virtual machines in cloud infrastructures, standard personal computers, or small-scale micro data centers.
