This Insights article was contributed by Mike Orr, Director of IT Services & Operations at Modoc Research Services
“For all the talk of machine learning and AI’s potential in the enterprise, many firms aren’t yet equipped to take advantage of it fully,” was the lead sentence of the April 27 Wall Street Journal article titled, "Machine Learning at Scale Remains Elusive for Many Firms."
With all the documented applications and apparent return on investment for businesses, why is this the case? The answer is that the move to AI requires a company to adopt a different computing model from the one that has evolved over the last 60 years. It is one in which the results mimic the interaction of a human with a changing environment.
With AI, you build the equivalent of a human expert in a particular domain. Like a human, this expert needs initial instruction, confirmation of knowledge, continuing education, and an acceptance by its users it may not always have the correct and/or expected answer.
Moving from the world of traditional sequential computing to AI is a journey or a climb up a ladder to a new level of computing.
The ladder from traditional computing to AI has four sequential rungs: Data, Analytics, Machine Learning (ML), and Natural Language Processing (NLP). Successfully climbing this ladder will provide the equivalent of a human expert in a specific domain.
For instance, Apple’s Siri personal assistant, Google’s Waymo autonomous car, and IBM’s Watson for Oncology, all have been developed, trained, interact with humans, and continue to learn based on their continued use. Each rung of the ladder can provide value to a company, and not all solutions have a requirement for true AI but many can advance through the stages in an orderly manner over time, gaining value as they evolve.
Like humans, an AI solution starts learning with data input. As infants, humans begin learning through gathering data via our senses. Computers must be provided data in a digital format. This may be textual data, digitalized images, smells, sounds, etc., or tactile information.
Preferably, the initial data needs to be credible and extremely consistent in format and content. As it begins processing the data, the computer must “assume’ it has correct and viable data to learn from. It knows no better. Thus, if an image recognition solution is being trained to recognize cows, and pictures of alligators are mixed in the training data, this “bad” data will affect what the solution thinks a cow “looks like.” As a company addresses the first rung of the AI ladder, it must ensure the initial training data is appropriate for the solution and of high quality.
Computer analytics allow a company to discover, interpret and communicate patterns in data. Most companies have data in multiple locations using multiple storage technologies. In addition, they have access to external public data.
Analytics provides a means of bringing multiple data sources together in one place and understanding significant patterns or connections that may exist. For example, a trucking company may be able to optimize its truck fleet scheduling through analyzing maintenance records and schedules, driver performance records, committed delivery schedules, and external data from the Weather Channel and fuel prices.
Analytics provides a means of bringing this data together and generating an optimum solution.
Machine Learning is focused on providing patterns of data to the solution and “teaching,” or adjusting, it to reach the preferred solution. The general technique is that, provided enough patterns and outcomes, the solution will have a context upon which to process an input pattern that it has never been exposed to.
The IBM Jeopardy! solution, which crushed its human opponents in the game show in 2012, was trained on thousands of previous Jeopardy! questions and how to find the correct answers in Wikipedia and other data sources. It was designed to be an expert on answering Jeopardy! questions and it succeeded through learning from thousands of questions and how to find their answers.
For this stage of the ladder, a company must dedicate its own Subject Matter Experts (SMEs) to the training of the solution. Only an expert can train a new expert. Many companies find this difficult to do, as their SMEs are busy providing support to other staff, briefing executive management, creating new products, identifying new markets, interacting with customers, and other activities a company expert is required to perform.
A technology services company can create the technical infrastructure for machine learning, but the understanding of what a company expert would do in a given situation must come from within. This investment is critical, though, as the computer student will only be as “smart” as the investment that has been made in its education.
NLP allows a solution to interact with humans using our communication methods of speech, written language, and/or visualization. This is in contract to a computers traditional communication method of a keyboard and screen of some type.
This rung requires a company to teach the solution the language of the domain and how humans utilize the language. This can be an extremely challenging task in some cases. Why does Siri get confused sometimes? Because natural language, which we take for granted, can be confusing and very contextual. What does it mean to be “cool”? Do noses really run and feet smell?
These are examples of the subtle contextual challenges that are continually faced by NLP programmers and trainers. Of course, there are entire vocabularies that may need to be learned for AI solutions concerning specific areas such as healthcare, finance or law.
Developing AI solutions involves a different set of company skills and resources than traditional computing applications. In some cases, the non-IT staff may be more involved than the technical staff. This new solution development and maintenance paradigm is difficult for some companies to understand and implement.
AI solutions are still in their infancy. As with any successful new technology, its implementation will become easier over time with new tools and shared experiences across industries and customer sets.
Diane Durance, MPA, is director of UNC Wilmington's Center for Innovation and Entrepreneurship (CIE). The CIE is a resource for the start-up and early-stage business community to help diversify the local economy with innovative solutions. For more information, visit www.uncw.edu/cie.
Jessica Maurer - Dec 11, 2018
Cece Nunn - Dec 12, 2018
Christina Haley O'Neal - Dec 11, 2018
Johanna Cano - Dec 12, 2018
Christina Haley O'Neal - Dec 10, 2018