Technology advancement offers an estimated $2 trillion in value for Supply Chain Management and Manufacturing. Attaining that value, however, will require separating the science from the science fiction surrounding AI.
Future visions of AI, both science and science fiction, depict robots in human form. Depending on your newsfeed, this could be a nightmarish outcome of an “accidental” creation as Elon Musk forebodes, or, as the humanoid robot Sophia self-describes, a contributor to mankind’s future that is built on human values such as compassion and kindness— even as she eerily calls out Musk in her debut “interview” at Saudi Arabia’s Future Investment Initiative.
Both scenarios are something to contemplate for the not-so-distant future, but data science experts including Brian McClendon, former VP at Google and co-creator of Google Earth, have a more realistic perspective. “What most people think of as AI, artificial intelligence, is still some ways out,” he says. True artificial intelligence is the point where machines can, literally, teach and think for themselves. What is currently billed as AI is, in truth, machine learning, a building block of AI.
Machine learning may not be nearly as sexy as a humanoid robot, but the value that ML offers for the nearer term, for existing business use cases is. Machine learning technology could yield an estimated value of $3.5 to $5.8 trillion for the global economy, according to this latest report by McKinsey. The financial impact will vary, notes the research, in the value for different roles and industries. The two biggest winners for their use cases are Marketing and Sales and Supply-Chain Management and Manufacturing. Supply Chain Management and Manufacturing is expected to gain $1.2 to $2.0 trillion in value from new technology. How?
McKinsey’s research focused explicitly on deep neural networks, networks that simulate connected “neural units,” mirroring human brain structure. It sounds very Westworld, but the applications for these deep learning systems are based on 400 real world use cases — existing business problems — within 19 industries and nine job roles. Use cases were not the stuff of an Arthur C. Clarke novel, but functions such as generating a route for vehicles to optimize time and fuel usage, or creating customer segmentation and predicting inventory needs.
Existing analytics processes already target many of these functions. Two-thirds of the predicted value in the report centered on use of machine learning processing to add an incremental lift to existing supply chain and manufacturing analytic techniques such as spend analytics, product development cycle optimization, inventory optimization, yield optimization, and risk modeling.
ML is particularly valuable when it comes to forecasting demand. By forecasting based on underlying causal market drivers rather than prior outcomes, accuracy can be improved by 10 to 20 percent, potentially reducing inventory costs by 5 percent, and increasing revenue by 2 to 3 percent. For example, pharmaceutical demand can often be hyper-regional, such as a flu epidemic. The technology can help predict these demands and relevant health trends, informing inventory levels and reducing spoilage. This application of deep ML offers potential to raise sales by 5 to 10 percent, says McKinsey analysts, while lowering costs.
For manufacturers, deep learning can also improve not just the modeling for risk, but reduce risk itself. Consider the systems used for monitoring temperature or vibration to forecast when machine parts may be likely to fail. With advanced technology, high dimensional data such as audio and visual information can be layered into the mix, making the prediction more accurate, as well as the timing of preventative interventions that reduce manufacturing and supply chain disruption.
At the demand end of the supply chain, next generation Sales and Marketing functions include informing pricing and promotions, sales forecasts, recommendation engines such as “others who buy” suggestions, and segmentation and targeting of marketing messages. The McKinsey report estimates that using customer data to personalize promotions can lead to a 1 to 2 percent increase in incremental sales for brick-and-mortar retailers alone.
Regardless which link in the supply chain, most of the value to be gained from ML is in operational efficiency and smarter decisions. And, unlike concept robots with simulated AI like Sophia, realizing the benefit from machine learning is near-term. However, a few barriers are holding us back from a smarter supply chain.
Hype, the first barrier on the trail to the AI frontier
While discussion of AI is everywhere, the technology is evolving. McKinsey research found that even among technically advanced firms, only about 20 percent are using one or more deep machine learning technologies in a core business process or at any scale. As companies navigate the new tech conversation, the current state of AI is analytics enhanced by machine learning.
Machine learning is computer programming that accesses data to deliver rule-based answers from a set of decision making patterns. It is a building block for AI.
Machine learning is to AI as repetition and rules are to human learning, whether we are learning language, math, or how to play a new game. This rote learning is one facet of how our brains work. But humans can also apply experiences to learn, create new ideas, and self-modify. We are capable of true cognitive learning where machines are not — yet.
Machine learning offers tremendous potential value, it is required for AI, a link in the chain toward machines that are capable of self-teaching. It may not seem as exciting as the science fiction, but machine learning can solve previously unsolvable problems. Here’s the reality filter you need to navigate new technology conversations.
Human intelligence still rules — and will for decades to come
Two-thirds of business use cases where machine learning offers value have existing analytics processes. Achieving added benefit — and hopes for ROI — requires a “training” set of data of labeled examples. Further, that initial labeling must be performed by humans who program the algorithm, and then “teach” the machine what the goal is and what to look for in the data.
The next step in the evolution is deep learning, as in the McKinsey report, or semi-supervised machine learning. Still, methods such as reinforcement learning and one-shot learning that reduce the human trainer effort are future capabilities. These capabilities will begin to close the gap toward cognitive, unsupervised learning, or full AI.
Access to user data will get increasingly difficult with data protection compliance
If your CIO is still recovering from the effort to reach compliance with the European Union’s General Data Protection Regulation (GDPR), this next generation of technology won’t make life easier. For one thing, access to and use of personal data will be more limited. But, a more complicated GDPR mandate relates to quantifying the use of that data. Per GDPR regulation, automated individual decision making— algorithms that make decisions based on user-level predictors—must allow for the user’s right to an explanation for decisions made by machines. With complex algorithms, there isn’t an easy explanation — or easy compliance.
By its nature, deep learning is useful for large and complex analysis. The McKinsey report highlights the challenges of this “last mile” link between machine’s complicated output and human’s comprehension, where the results must be clear and actionable. These regulatory and human challenges will also hinder adoption of more advanced machine learning, and full AI, in industries that require product certifications with clearly defined rules and choice criteria, including healthcare, automotive, chemicals, and aerospace industries.
To err is human — and machine
Perhaps the eeriest aspect of Sophia’s robotic “interview” is not her simulated human expressions and perfect countenance, it’s her calculating, mechanical superiority. “I know humans are smart and very programmable,” she says. “I want to use my artificial intelligence to help them live a better life.”
What, or who, defines “better?” Despite this scary display, cognitive, self-aware artificial intelligence does not exist yet. For one thing, computers are not capable of “transfer learning,” or our very human ability to apply experiences from one set of circumstances to another. Unlike our adaptable brains, machines still must be trained anew even for similar use cases.
ML systems are also prone to errors from “overfitting” where a learned model too closely matches random features in a set of training data — generalizing for us non-robots — or “underfitting” where the model fails to capture all the relevant information. All of this sounds remarkably similar to human errors, including our shared weakness for bias, which results from decisions based on a limited set of data.
When it comes to AI, build your intelligence first
Beyond the creepy valley, as Sophia explains, “the concept that if robots become too humanistic, they become creepy,” cognitive AI is an exciting new frontier. But machine learning is accessible now and offers tremendous potential for business. The most important way to prepare for the future is to improve your knowledge on emerging tech.
For forward-thinking organizations, this requires evaluating what use cases and technology will be a fit for your business strategy and offer actual ROI. Now is also a good time to understand how the coming “Fourth Industrial Revolution” of AI, robotic process automation, and machine learning will impact your role and industry, and the “human intelligence” you will need to hire in the future.
Perhaps most importantly, you should learn enough about these technologies to ask the right questions from the vendors selling solutions. Questions that focus on your very real business needs. This intelligence will save you the cost of buying both the wrong expensive solution and the hype.