Presented By O'Reilly and Cloudera
Make Data Work
September 26–27, 2016: Training
September 27–29, 2016: Tutorials & Conference
New York, NY

Removing complexity from scalable machine learning

Martin Wicke (Google)
5:25pm–6:05pm Wednesday, 09/28/2016
Data science & advanced analytics
Location: Hall 1C Level: Intermediate
Average rating: ***..
(3.50, 2 ratings)

Prerequisite knowledge

  • A fundamental understanding of machine learning and deep learning
  • What you'll learn

  • Explore emerging tools that can dramatically impact the feasibility of the most ambitious projects
  • Description

    Much of the success of deep learning in recent years can be attributed to scale—bigger datasets and more computing power—but scale can quickly become a problem. Distributed, asynchronous computing in heterogenous environments is complex, hard to debug, and hard to profile and optimize. Martin Wicke demonstrates how to automate or abstract away such complexity, using TensorFlow as an example. Martin covers the sources of complexity for large-scale machine-learning systems, explains how to mitigate such complexity, and touches upon the future avenues for this work, where, unsurprisingly, machine learning will be used to understand and improve machine learning.

    Topics include:

    • Heterogeneous systems: In most cases, we have to contend with at least CPUs and GPUs, but for very large problems, FPGAs and custom hardware are increasingly used.
    • Code complexity: As AI models become more complex, traditional software engineering issues creep into the everyday lives of researchers. Problems with code reuse, proper abstraction, and data isolation become increasingly important.
    • Model complexity: The inner workings of complex models are impossible to understand without custom tooling. Custom visualization and debugging tools are increasingly necessary, but specific approaches can help understand certain classes of models. (For instance, Deepdream is based on a tool to better understand how convolutional neural networks work.)
    Photo of Martin Wicke

    Martin Wicke


    Martin Wicke is a software engineer at Google working on making sure that TensorFlow is a thriving open source project. Previously, Martin worked in a number of startups and did research on computer graphics at Berkeley and Stanford.