Presented By O’Reilly and Intel AI
Put AI to work
8-9 Oct 2018: Training
9-11 Oct 2018: Tutorials & Conference
London, UK

Scaling deep learning on AWS using C5 instances with MXNet, TensorFlow, and BigDL: From the edge to the cloud

GAURAV KAUL (Amazon Web Services), Suneel Marthi (Amazon Web Services), Grigori Fursin (dividiti)
13:45–14:25 Wednesday, 10 October 2018
Location: Westminster Suite
Secondary topics:  Deep Learning tools, Edge computing and Hardware, Platforms and infrastructure

What you'll learn

  • Learn how to optimize cost, performance, and convergence for your deep learning models, using CPU spot instances in AWS

Description

Gaurav Kaul and Suneel Marthi share trade-offs and design choices that are applicable to deep learning models when training in the cloud, specifically focusing on performance, efficiency and reproducibility. They then demonstrate how to optimize for performance and cost using CPU spot instances in AWS. Grigori Fursin will then show how to automate and crowdsource this complex process of optimization, autotuning and co-design of efficient software and hardware using the open-source Collective Knowledge framework (CK). He will also present an open repository of reusable CK workflows and components (models, data sets, frameworks) with contributions from the 1st reproducible optimization tournament for deep learning (ACM ReQuEST) showing some 50x speedups with negligible accuracy loss for 8-bit inference on Intel Xeon processors in the Amazon AWS cloud.

Photo of GAURAV KAUL

GAURAV KAUL

Amazon Web Services

Gaurav Kaul is an architect at AWS in the UK.

Photo of Suneel Marthi

Suneel Marthi

Amazon Web Services

Suneel Marthi is a principal technologist at Amazon Web Services. Suneel is a member of the Apache Software Foundation and is a PMC member on Apache OpenNLP, Apache Mahout, and Apache Streams. He has given talks at the Hadoop Summit, Apache Big Data, Flink Forward, Berlin Buzzwords, and Big Data Tech Warsaw.

Photo of Grigori Fursin

Grigori Fursin

dividiti

Grigori Fursin is the CTO of dividiti and a founding member of the ACM taskforce on reproducibility of the ML, AI, and systems research. Previously, he was a tenured scientist at Inria and the head of the Optimization Group at the Intel Exascale Lab. Grigori is the architect of the Collective Knowledge technology used by a growing number of universities and Fortune 50 companies to enable automatic, collaborative, and reproducible development, optimization, and codesign of efficient software and hardware for emerging ML and AI workloads in terms of speed, accuracy, energy, size, and costs. He holds a PhD in computer science from the University of Edinburgh.