Schedule: Hardware sessions

Success with DL requires more than just TensorFlow or PyTorch. Angela Wu, Sidney Wijngaarde, Shiyuan Zhu, and Vishnu Mohan detail practical problems faced by practitioners and the software tools and techniques you'll need to address the problems, including data prep, GPU scheduling, hyperparameter tuning, distributed training, metrics management, deployment, mobile and edge optimization, and more. Read more.

Long training times are the single biggest factor slowing down innovation in deep learning. Today's common approach of scaling large workloads out over many small processors is inefficient and requires extensive model tuning. Urs Köster explains why with increasing model and dataset sizes, new ideas are needed to reduce training times. Read more.

Joseph Spisak and Hao Lu lead a deep dive into how PyTorch is being used to help accelerate the path from novel research to large-scale production deployment in computer vision, natural language processing, and machine translation at Facebook. Read more.

Over the last few years, convolutional neural networks (CNNs) have risen in popularity, especially in the area of computer vision. However, CNNs are by nature computationally and memory intensive, making them challenging to deploy on a mobile device. Siddha Ganju and Meher Kasam examine optimizing deep neural nets to run efficiently on mobile devices. Read more.

5G promises to change our lives in a big way. Mazin Gilbert provides a technical- and market-landscape overview of how AI creates the 5G world, highlighting how recent developments in AI help accelerate widespread adoption of 5G-based applications for consumers and enterprises. He explores the roles of open source and open platforms as key ingredients of this 5G AI transformation. Read more.

Dejan Milojicic examines a software stack designed for the special-purpose machine learning accelerator. The software stack improves usability and programmability of the accelerator, making it accessible from common machine learning frameworks. The software toolchain also exposes the intricacies of the parallelism of the accelerator while hiding its complexities. Read more.