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 long and winding road to AI: Lessons from implementing cognitive AI

Rupert Steffner (WUNDER)
4:50pm–5:30pm Wednesday, May 2, 2018
Implementing AI
Location: Sutton North/Center

Who is this presentation for?

  • AI engineers, AI architects, and deep learning engineers

Prerequisite knowledge

  • A basic understanding of AI implementations and multidiscipline and model thinking

What you'll learn

  • Explore lessons WUNDER learned implementing cognitive AI applications to help consumers find the products they love

Description

WUNDER is one of the few companies that built real-world cognitive AI solutions in 2017. But why did it take long to bring AI applications into the real world? And why do shortcuts fail so often, as indicated by the 70% failure rate of chatbots? The road to real-world AI is long and winding. All we’ve heard from reputable experts turned out to be true.

Peter Norvig was right: AI needs better data. In fact, it needed a whole new data type, which we call cognified product data. It gives products a human-level personality and makes products matchable with human personalities. This resulted in a tremendous effort of building a very special blend of product data preparation through a 15-layered product graph warehouse. It has been a hard lesson to learn that shaping AI-ready data goes far beyond data preparation for business intelligence. And Mike Loukides was right: AI needs a new UX. AI engineers rarely ask what the experience is users want in being assisted. It can be frustrating to have a perfect solution that doesn’t fit the user’s problem. AI is about context.

Rupert Steffner highlights lessons learned implementing cognitive AI applications to help consumers find the products they love. Rupert explains how WUNDER really had to understand its users and design the experience as close as possible to their learned shopping behavior to make the trade-off as minimal as possible. The company found the right metaphor in a consumer entering a shop in the real world and built a playful experience around this context by integrating game design elements for interaction and navigation. In order to get the right context, WUNDER had to build up its own sensing inventory and make sure that the sensors would be operated the right way. This AI system is sensing preferences, activities, styles, and, most important, emotions. Emotions are operated in the human’s limbic brain, which has no capacity for language or words. Thus, the sensing inventory had to work with images and videos to ensure it could be interpreted by the consumer’s brain.

For the AI execution framework, WUNDER had to decide whether to switch to Ray or to continue with its own real-time streaming analytics framework. The company decided to go with its in-house framework (now in its third generation) as it offered much more flexibility to design and execute the sense-infer-act-learn paradigm for the cognitive AI application. It’s hard to find the few developers capable of building AI applications based on in-memory computing and data locality to get the execution done in milliseconds. Since nobody had ever designed real-time streaming topologies the way WUNDER needed it, it took some time to align the streaming topology with the conceptual AI workflow and interaction and learning loops.

WUNDER also learned it had to look at learning more broadly. The company explored concepts like deep symbolic reinforcement learning to combine pure pattern recognition with prior domain symbol knowledge and tried to bring pattern recognition with higher order reasoning into play. It then found ways to work with less data and training using testing approaches like Josh Tenenbaum’s one-shot learning. It was also highly influenced by insights from neuroscience, taking DeepMind’s “replay” and “imagination” concepts as example. Finally, WUNDER was more conscious of what parts of the system are traits or states and which undergo which speed of learning. One key takeaway is that excluding the system’s underlying value system from fast learning is a fairly good idea to prevent what happened to Microsoft’s Tay. Learning in AI is widespread and goes far beyond machine learning. Will future machine learning engineers need to bring in neuroscience and cognitive psychology skills? The good news is that we all have just begun to scratch the surface.

However, to manage its new technological AI platform, WUNDER needed a native AI architecture, which the company created as a logical architecture with stacks (preparation, learning, execution, UX, and exchange) and layers—and did the same with the technology architecture, creating a hybrid cloud and on-premises architecture. This allowed the company to improve piece by piece whenever it was needed or new ideas came in. But even more importantly, the architecture allowed WUNDER to switch from building the platform to feature-driven development.

Photo of Rupert Steffner

Rupert Steffner

WUNDER

Rupert Steffner is the founder of WUNDER, a cognitive AI startup that is helping consumers find the products they love. Rupert has over 25 years of experience in designing and implementing highly sophisticated technical and business solutions, with a focus on customer-centric marketing. Previously, Rupert was chief platform architect of Otto Group’s new business intelligence platform BRAIN and head of BI at Groupon EMEA and APAC. He also served as business intelligence leader for several European and US companies in the ecommerce, retail, finance, and telco industries. He holds an MBA from WU Vienna and was head of the Marketing Department at the University of Applied Sciences in Salzburg.