Report from ai.now: Giving Vision to AI
During a presentation and discussion at the ai.now Workshop in July, David Dittmann, director of business intelligence and analytics services for Procter & Gamble, explained how to “Give Vision to AI” with Pranay Agrawal, co-founder and chief executive officer of event sponsor Fractal Analytics.
By 2020, 80% of the data created will be unstructured, Dittman explained, and that offers immense opportunity to analyze images and videos — with new data bringing new signals.
To kick off the discussion, Dittmann shared the story of Makoto Koike, who began using deep learning at his family’s cucumber farm in Japan. When he started helping at the farm, Koike — a former embedded systems designer for the automobile industry was amazed by the amount of work it takes to sort cucumbers by size, shape, color and other attributes.
Since each of the nine different varieties of cucumber has a different color, shape, quality and freshness, it was taking Koike's mother up to eight hours per day to sort during peak season. Knowing there must be a better way, Koike got the idea to implement deep learning to sort the cucumbers, saving hours per day.
Using deep learning for image recognition lets a computer learn from a training data set what the important "features" of the images are. By using a hierarchy of numerous artificial neurons, deep learning can automatically classify images with a high degree of accuracy. Often, speed and accuracy can exceed that of the human eye.
However, critical to an AI deployment is preparation for failure. “No matter what your vision is, your data is key,” says Dittmann. Early AI stirs excitement, but if companies jump in too soon, the data needed to truly drive value sometimes isn’t there. Once that data is in place, machine learning will begin to bloom, and then deep learning breakthroughs will drive an AI boom. In the meantime, companies should be prepared for a lot of learning curves — and a series of failures.
Aside from internal failures, there are other issues with AI that the organization must be prepared to overcome. It’s very difficult to understand how a “deep learning” model is making decisions (is it identifying a cat’s face, or is it identifying the colored background?). Scored training data is critical. Although it requires a lot of time and teaching, once the deep learning model is identified and corrected, the reward is abundant.
Artificial intelligence has great potential throughout a number of industries, and opportunities are often endless. For example, in a medical field the ability to automate reports based on diagnosis could save time for both physicians and patients. In current trials in India, AI allows doctors to diagnose cases of tuberculosis early on, potentially saving lives. In the insurance industry, AI could help streamline assessments, automating the claims process. Even the Super Bowl is able to leverage image recognition software and AI to run analysis on marketing program impressions.
The underlying theme throughout all these use cases? It’s all about the data. The ability to take unstructured data and make it structured will help close the informational loop. Structured data output allows for further analysis. In stores, the use of structured data and analysis lets the shelf work harder and, therefore, can simplify planograms.
The big question in an organization? Where to start and how to make this vision work. First, companies must identify the business situation of interest. Second, they should establish a source of image information and continually check for visual insights. Making vision work requires evaluation and prioritization of various use cases. Lastly, companies must pilot and learn.
“A beautiful world of possibility awaits for AI,” said Agrawal. “There is tremendous power in AI to make this world a better, more meaningful place for all of us.”