Presented By O’Reilly and Intel AI
Put AI to Work
April 29-30, 2018: Training
April 30-May 2, 2018: Tutorials & Conference
New York, NY

The vital role of failure in machine learning

Scott Weller (SessionM)
11:55am–12:35pm Wednesday, May 2, 2018
AI Business Summit, AI in the Enterprise
Location: Sutton South
Average rating: **...
(2.33, 3 ratings)

Who is this presentation for?

  • C-suite executives, decision makers, managers, and technologists

Prerequisite knowledge

  • A basic understanding of machine learning
  • Familiarity with how AI algorithms are trained (useful but not required)

What you'll learn

  • Understand where we are with AI today, how to set realistic expectations when it comes to AI deployments, and examples of good and bad AI deployments in the wild


Solving any problem involves calculations. Whether working to determine the effectiveness of a new medicine or trying to predict the success of a Mars rover, innovators look at all the factors in play and use them to determine the outcome before deploying a new solution.

This process used to rely on manual calculations to simulate outcomes, followed by trial and error. However, affordable pricing and wide adoption of cloud computing has made machine learning widely accessible at scale. Today, with AI, these calculations can happen at a pace previously impossible. Simulation as a practice is more accessible to everyone, and the ability to simulate the outcome of a variety of factors and features has dramatically cut down on the “error” side of “trial and error” when it comes time to try a solution. In a way, the world is now like a game.

But with games in mind, there’s a flaw in how we train AI systems today. Algorithms are currently built off of success, learning from wins and coming from a place of positive reinforcement. In a game, however, you learn by failing—rapidly and repeatedly. You might “die” hundreds of times in a video game, but this process of aggressive failure is a necessary part of training you to succeed.

Similarly, AI comes with risk. We need to expect failures straight out of the box; we can’t expect algorithms that are trained by success to be successful immediately. The best algorithms will fail first, sometimes in ways that will make those that implemented them think they’re going to get fired. In business, implementing AI could mean losing money during the first quarter of deployment. Enterprises need to ask themselves, “Are we resilient enough to weather failures before seeing success? What are we risking by using AI?”

Scott Weller explores the role of failure in machine learning, explaining where we are with AI today, how to set realistic expectations when it comes to AI deployments, and examples of good and bad AI deployments in the wild.

Photo of Scott Weller

Scott Weller


Scott Weller is cofounder and CTO of SessionM. Scott has over 18 years of development, operational, and leadership experience turning ideas into reality and leading technology teams through the challenges of early-stage growth. Previously, he was vice president of product and technology for Scientific Games (SGMS), where he oversaw the development and integration of interactive technologies into MDI’s products and services; vice president of product and technology at GameLogic (acquired by Scientific Games in 2010); cofounder and GM of; principal software engineer at Terra/Lycos, where he spent several years innovating data and advertising platform technologies; and senior software engineer at (acquired by Lycos in 1999). At the age of 16, he joined a team of eight motivated geeks to help build the country’s first internet service provider, later acquired by Conversant Communications. Scott holds a BS in computer science from the University of Rhode Island.