Unshattering the mirror: Defragmenting the deep learning ecosystem

Evan Sparks (Determined AI)

2:35pm–3:15pm Wednesday, September 11, 2019

Location: 230 A

Implementing AI

Secondary topics: Deep Learning tools

Who is this presentation for?

Directors of AI, VPs of engineering, deep learning (DL) developers, ML engineers, and data scientists

Level

Intermediate

Description

Companies like Google, Facebook, Amazon, and Microsoft have comprehensive internal tooling that enables their developers to collaborate and build end-to-end AI applications. These platforms allow developers to be insanely productive and deliver new AI features and applications orders of magnitude faster than the rest of the industry. Meanwhile, developers at every other organization on the planet are left piecing this infrastructure together with a combination of highly specialized point solutions, legacy systems not designed for modern AI workflows, and half-baked open source projects. These tools lack standard interfaces and file formats, and they’re often incompatible in surprising ways.

Evan Sparks details the deficiencies with existing reference architectures for AI development infrastructure and the opportunities for end-to-end system design in AI development with deep dives into two examples: there’s orders of magnitude improvement in training performance and convergence to better models by integrating cluster resource management and fine-grained scheduling with hyperparameter optimization, and workload-aware checkpointing enables seamless and rapid fault tolerance, auto-scaling, and rapid collaboration.

Prerequisite knowledge

A working knowledge of the model development process
A high-level understanding of the existing tooling in the DL ecosystem for tools like model development, hyperparameter optimization, model compression, and training cluster management

What you'll learn

Gain a better understanding of the gaps in current publicly available AI development infrastructure and a good sense of what you should demand from this infrastructure

Evan Sparks

Determined AI

Evan Sparks is a cofounder and CEO of Determined AI, a software company that makes machine learning engineers and data scientists fantastically more productive. Previously, Evan worked in quantitative finance and web intelligence. He holds a PhD in computer science from the University of California, Berkeley, where, as a member of the AMPLab, he contributed to the design and implementation of much of the large-scale machine learning ecosystem around Apache Spark, including MLlib and KeystoneML. He also holds an AB in computer science from Dartmouth College.