Presented By O’Reilly and Intel Nervana
Put AI to work
September 17-18, 2017: Training
September 18-20, 2017: Tutorials & Conference
San Francisco, CA

Backing off toward simplicity: Understanding the limits of deep learning

Stephen Merity (Salesforce Research)
11:55am–12:35pm Tuesday, September 19, 2017
Implementing AI
Location: Imperial A
Secondary topics:  Deep learning, Technical best practices
Average rating: ****.
(4.62, 8 ratings)

Prerequisite Knowledge

  • A general understanding of neural networks, particularly recurrent neural networks

What you'll learn

  • Discover which tasks benefit most from deep learning
  • Understand the common trade-offs when research-focused deep learning architectures move toward production
  • Explore successful use cases of production systems that draw on but fundamentally modify existing state-of-the-art research


Deep learning is used broadly at the forefront of research, achieving state-of-the-art results across a variety of domains. However, that doesn’t mean it’s a fit for all tasks—especially when the constraints of production are considered. While in some cases, deep learning can be applied without thought, most domains require understanding the task and the trade-offs involved when crafting a specific solution, especially when the system is designed with production in mind.

Exploring successes in both research and production, Stephen Merity investigates what tasks deep learning excels at, what tasks trigger a failure mode, and where current research is looking to remedy the situation. By pulling apart specific examples, such as Google’s Neural Machine Translation architecture or Salesforce Research’s quasi-recurrent neural network, Stephen analyzes the trade-offs made when stepping away from research toward production systems, noting when deep learning is likely the wrong tool of choice, especially when factoring in real-world restrictions, such as training a custom model for each customer or tackling vast datasets.

Photo of Stephen Merity

Stephen Merity

Salesforce Research

Stephen Merity is a senior research scientist at Salesforce Research (formerly MetaMind), where he works on researching and implementing deep learning models for vision and text, with a focus on memory networks and neural attention mechanisms for computer vision and natural language processing tasks. Previously, Stephen worked on big data at Common Crawl, data analytics at, and online education at Grok Learning. Stephen holds a master’s degree in computational science and engineering from Harvard University and a bachelor of information technology from the University of Sydney.