Meet Your New Assistant: How AI and Digital Twins Are Enhancing the Way We Work
September 14, 2017
GE Reports Canada
We tend to view AI development as a high-stakes game of human-versus-machine, but with tools like Digital Twins, it’s collaboration that wins the day.
The impact of artificial Intelligence on industrial production is well underway and still growing, yet the topic sometimes produces as much anxiety as it does excitement, especially for those who fear that their jobs may become automated.
“There are two main fears with AI – first that there will be a loss of jobs, and second, the black box nature of machine learning. These are the main challenges we’re trying to solve with industrial AI,” says Achalesh Pandey, Technology Leader for Artificial Intelligence & Learning Systems at GE Global Research.
While the dire outlook on AI’s impact is prevalent, there exists an alternative scenario that is already being put into practice: collaboration. Today, there are many cases where humans and machines are combining their efforts and learning from each other. This new form of teamwork results in smarter machines and more effective decision-making, particularly in large-scale industrial contexts.
Building trust for AI
Gabriel Duford, SVP Development and Technology and Co-founder at Element AI in Montreal, points to supply chain management, especially Retail Planning, as an example where AI is having a disruptive impact. Until recently, retail planners operated intuitively, using a combination of gut instinct and experience to decide what merchandise to order. Those decisions can now be made more accurately using AI tools that can simultaneously consider many more variables than a human would ever be able to.
But despite the obvious benefits, acceptance of these new recommendations doesn’t come easily. “There’s no way a machine is going to tell me what to put on the shelf,” an experienced manager might say.
Generally, trust in an AI recommendations’ takes time, increasing gradually as humans start to see benefits of working with this new technology. Pandey explains the approach at GE, where collaboration with domain experts is prioritized.
Rather than coming up with an AI solution independently, experts are consulted and involved in the process of creating the AI solution. Once the solution is implemented, it is then much more likely to be relevant and accepted. Another change management strategy involves identifying members of a team who might be more likely to embrace new ideas, and letting them lead the transition towards working with AI. This fusion of human domain knowledge and machine learning is key.
Duford echoes this notion of an open mindset, noting that those with a flexible attitude will likely have a much easier time adapting to working with AI: “We have to accept that we may not know everything, and the AI may know better. People that are more humble will succeed faster at leveraging an AI to do a better quality and more efficient job.”
Another way that humans and machines are developing stronger bonds involves the application of reinforcement learning. It’s not a new concept, but has gained traction with recent advances in deep learning.
Psychologist Edward Thorndike was one of the first to document reinforcement learning in the early 20th century, carrying out experiments with cats. Tasked to escape from a box, the cats could only get out by pressing a lever. At first, they succeeded only by accident, and then with increasing speed as they made the association between the lever and the exit.
Eric Laufer, an Applied Research Scientist in Artificial Intelligence at Element AI, describes how this empirical approach is applied in a machine learning context: “The machine isn’t told what the best actions are. It simply has to try actions for itself, and then observe the results, making adjustments in order to maximize optimal results. It’s like a trial and error framework for interacting with the world.”
This approach to problem solving is being used to great effect at Kindred AI, a Canadian robotics startup that pairs reinforcement learning with human intervention, aiming to amplify learning capabilities for both parties.
They realized that the best way to teach robots to carry out difficult dexterous tasks, such as selecting and picking up different types of objects, was to pair them with human “pilots” using VR headsets and motion-tracking controllers. Machine learning algorithms help robots carry out a desired outcome, such as picking up an object, but if the probability of success is low, the bot can call for human help. In this scenario, the AI’s capabilities are enhanced by learning from the actions of a human, while the human’s job is made easier by eliminating the bulk of repetitive manual labour, for example in a retail warehouse context.
Human-machine dialogue – meet your Digital Twin
This notion of a continual exchange of knowledge from AI to humans is one that is already well underway at GE, where the concept of Digital Twins is facilitating a new type of communication—one that allows humans to anticipate future mechanical issues and maximize efficiencies.
While the notion of building a model for monitoring and testing is not new—NASA was doing it in the early days of space travel—the recent addition of deep learning to the equation has made the process more sophisticated. Through GE’s Cloud-based Predix operating system, humans can literally converse with large-scale machinery such as wind turbines, jet engines, or locomotives.
Pandey compares the phenomenon to the first digital shift, when paper trails in offices were replaced with digital files. There was an initial confusion, followed by acceptance and adaptation when we discovered the convenience of sharing digital information vs. physical information.
With the ongoing digital industrial shift, however, he points out that the physical asset doesn’t completely disappear, but rather co-exists with its Digital Twin counterpart. When a physical asset, such as a jet engine for example, is affected by environmental conditions such as degradation, aging, macro or micro economic changes, its Digital Twin evolves as it adapts alongside it. The relationship between the physical asset and the Digital Twin is symbiotic, featuring a continuous exchange of knowledge and information.
An AI then consolidates and curates all of this data into relevant recommendations, allowing human domain experts to make much more sophisticated decisions than they would have otherwise.
GE Transportation is applying Digital Twins in the locomotive industry, and extending the concept to include the “Digital Thread”, which entails digitally tracking how a train is designed, configured, built, operated, and serviced.
The process of formulating the Digital Thread allows GE to track the health of every component of a train, in terms of its history, stresses, strains, performance, emissions, operating and environment. All of this information and ability to continually communicate with formerly non-sentient machinery is allowing humans to make better decisions, thereby optimizing the lifespan of their machine counterparts, and making them more efficient.
While we’re still in the early days of the digital-industrial revolution, it’s amazing how quickly our relationship with AI is evolving. While we used to think of learning as a transfer of knowledge from a teacher, book, or the internet, we’ll have to get more comfortable interacting with machines, as they become increasingly sophisticated through AI. As Pandey reminds us, “Nothing is fixed or constant—these things evolve with time.”