Schedule: Edge computing and Hardware sessions

Machine learning sits at the core of many essential products and services at Facebook. Yangqing Jia and Dmytro Dzhulgakov offer an overview of the hardware and software infrastructure that supports machine learning at global scale. Read more.

Nigel Toon explains how scaling IPUs will increase the productivity of machine intelligence researchers everywhere. Join in to explore what can we do and expect from the field with vastly more compute. Read more.

Gaurav Kaul, Grigori Fursin, and Suneel Marthi share trade-offs and design choices that are applicable to deep learning models when training in the cloud, specifically focusing on convergence and numerical stability, which are very important for autonomous driving and medical imaging. They then demonstrate how to optimize cost, performance, and convergence using CPU spot instances in AWS. Read more.

Shaoshan Liu explains how PerceptIn built the first FPGA-based computing system for autonomous driving. Read more.

Bruno Fernandez-Ruiz details a unified network that jointly performs various mission-critical tasks in real time on a mobile environment, within the context of driving. Along the way, he outlines the challenges that emerge when training a single mobile network for multiple tasks, such as object detection, object attributes recognition, classification, and tracking. Read more.

Kaz Sato offers an overview of ML Ops (DevOps for ML), sharing solutions and best practices for bringing ML into production service. You'll learn how to combine Apache Airflow, Kubeflow, and cloud services to build a data pipeline for continuous training and validation, version control, scalable serving, and ongoing monitoring and alerting. Read more.

Recent research has shown that training for quantization can lead to large gains in energy efficiency, and embedded runtime packages like TensorFlow Lite and Caffe2Go offer portability over a number of platforms. Cormac Brick asks, Why can't we have both performance and portability? Cormac explores industry challenges and details the progress needed to close the portability-performance gap. Read more.

SURFSara and Intel collaborated as part of the Intel Parallel Computing Center initiative to advance the state of large-scale neural network training on Intel Xeon CPU-based servers. Ananth Sankar, Valeriu Codreanu, Damian Podareanu, and Steve Smith share insights on several best-known methods for neural network training and present results from tests performed on Stanford's CheXNet project. Read more.

Machine learning and AI models now outperform humans on many tasks. However, sending sensor data up to the cloud and back is too slow for many apps and autonomous machines. Jameson Toole explains why developers seeking to provide seamless user experiences must now move their models down to devices on the edge, where they can run faster, at lower cost, and with greater privacy. Read more.

In recent years, we’ve seen a shift from traditional vision algorithms to deep neural network algorithms. While many companies expect to move to deep learning for some or all of their algorithms, they may have a significant investment in classical vision. Paul Brasnett explains how to express and adapt a classical vision algorithm to become a trainable DNN. Read more.

Zhipeng Huang explains how resource representation (RR) works with various intermediate representation (IR) technologies to help achieve the democratization of AI. Read more.