Lifelong Learning Machines to offer even more intelligent AI

DARPA has proposed the Lifelong Learning Machine while researching the development of machine learning technology. L2M would use previously learned skills to adapt to novel situations, something that would greatly benefit metalworkers and other manufacturers. Getty Images.

The industrial internet of things (IIoT) and artificial intelligence (AI) offer promising opportunities to improve manufacturing results, and that can cause metal formers and fabricators concern about job stability.

The thing to keep in mind is that people have a significant advantage over their machines. People learn from their experiences and build on previously acquired knowledge. Tool and die builders, operators, engineers, and designers all can draw on their own experiences and those of others. For now, computers can execute predefined tasks within given parameters, and machine learning is limited to isolated events. They don’t have the capability to learn from related events or learned experiences.

Remember how you learned mathematics? You started in elementary school by learning the four basic operations: addition, subtraction, multiplication, and division. As you advanced through the grades, teachers introduced you to fractions, decimals, sets and subsets, and fundamental algebra. Each new year’s instruction probably refreshed what you learned the year before, then built on what you learned by introducing new representations and functions, eventually leading to algebra, geometry, trigonometry, and possibly calculus.

And now as an adult, when your components fail on the shop floor, you likely follow some form of a decision tree, identifying the most likely causes, testing the process or equipment, analyzing the result, identifying causes and symptoms, and then implementing a remedy. If you can’t identify a likely cause, you look back to your experiences and ask others if they encountered similar problems before.

Computing technologies don’t work that way. While people retain memories of experiences that are immediately available to adapt their behavior to a dynamic set of circumstances, computers provide programmed and predictable responses to a given set of inputs. With responses locked in during development, computers are unable to respond to the huge amount of eventualities in information input in a manufacturing operation.

Even neural networks are unable to expand their repertoire of behaviors in response to new information. Neural networks are prone to “catastrophic forgetting,” which is the tendency of the network to forget previous learning when new and different information is introduced.

Even in the simplest application, a machine’s ability to “learn” is very limited. As autonomous systems are connected in a network, the introduction of new information increases and learning systems fail under today’s computing model.

The Defense Advanced Research Projects Agency, Arlington, Va., is researching the development of technologies that learn continuously. It has proposed a machine called the Lifelong Learning Machine (L2M), which would have the following fundamental capabilities:

• Learn continuously, even during execution of tasks
• Use previously learned skills to adapt to novel situations without forgetting learned tasks
• Understand and adapt to input signals within the context of a given goal and contextual influences
• Know when to learn and when not to learn
• Maintain safety and always execute correct behavior in a continuously changing system

None of these capabilities is trivial, and none can be readily developed with today’s technologies. Some advances have been made in teaching machines by doing, but that research is still in the early stages. Successful implementation of L2Ms in practice will first require they meet practical management mandates, including:

• Energy management – A continuous learning machine requires tremendous amount of energy.
• Safety – No system can be introduced without confidence that it will not risk the safety of employees, environment, or equipment.
• Success measurements – New systems must offer both financial and functional value in a forming operation.
• Integrity – L2M systems must truly be learning systems, rather than cleverly programmed facades.
• Human interaction – Systems must support and empower employees to grow in their knowledge and capabilities.

Industry has many years of research, development, and certification ahead before L2Ms are commercially viable and available. Because L2Ms require some changes in the basic nature of computing, researchers currently don’t understand the capabilities and limitations of the evolving technology.

As this technology continues to advance and improve, metal forming and fabrication companies must be at the top of their game to realize the opportunities and manage new information technologies, as well as new and evolving materials. The more you understand available and evolving technologies, the better you will be able to learn and profit from their capabilities and manage their limitations.