Benchmarks and How-Tos for Convolutional Neural Networks on HorovodRunner-Enabled Apache Spark Clusters

Explore benchmarks and implementation techniques for Convolutional Neural Networks using HorovodRunner on Apache Spark clusters in this 25-minute Databricks talk. Learn about the scaling efficiency of Horovod Runner for CNN-based tasks on both GPU and CPU clusters, and discover the first-time implementation of the Rectified Adam optimizer. Gain insights into cluster settings, distributed model retrieval, training time measurement, Rectified Adam usage, and factors affecting scaling efficiency. Understand how smaller companies can leverage distributed deep learning for competitive advantages against tech giants through fast iterations on accessible Spark clusters.

Introduction
Why a Distributed Deep Learning System
Data Parallelization
Benchmarks
HorovodRunner Demo
Benchmarking
Graph
Ubers
Applet
Other Models
Summary