Machine Learning Tailors Training to the Student
By Tobias Naegele
March 13, 2017
Share this page:
Most training and learning systems today follow the same basic model developed a century and a half ago when the British pioneered industrial-scale education to produce a literate working class. With one teacher and a large group of students, instructors must focus on the ones in the middle. While the weakest in the class may drop out, the most talented are slowed down and forced to be average. Machine learning could change all that.
Training systems today can capture individual performance characteristics and, with the help of data analytics, identify strengths and weaknesses so training can be tailored on the fly to match the specific needs and requirements of each individual student.
“You don’t want to present each learner with the same experience, with the same knowledge,” explains Luke DeVore, business development director at Design Interactive of Orlando, Fla., a human-centered design firm with a focus on training. “Some people have different starting points. They have different backgrounds that allow them to progress more quickly. Or they might move more slowly through certain sections.”
Daniel Serfaty, founder of the human-centered engineering firm Aptima and a pioneer in simulation science, says the ability to tailor individual training for each individual to the individual is the next great step in the evolution of one of training technology, made possible by the advent of affordable artificial intelligence (AI).
AI can be used first to analyze individual performance and later to tune training to trainees’ specific weaknesses. Eventually, he and others say, AI will enable everyone to have a personal learning partner to enhance on-the-job performance.
“As the gap between our capabilities and those of our [rivals] gets closer,” Serfaty told military training leaders in December, “I believe the last frontier in which America and its allies can still have an advantage over our adversaries, is human performance.” To achieve that, he added, America must “take learning to the next level.”
It’s well established that different people learn at different rates. Yet today’s training systems aren’t built with that in mind.
“We have to develop technologies that enable us to adapt and personalize the training so that we can tailor that training to the individual,” Serfaty said. This, he added, is the “low-hanging fruit” in the race to leverage machine learning in training technology.
Using machine learning to identify the weaknesses of human performance – and then help them overcome those weaknesses by adapting the performance intervention to the correct level of knowledge and skill – is more than an ironic twist. “Making sure weaknesses are identified early in the training is a much better way to ensure that training results are more effective and more efficient,” says Dr. Denise Rose Stevens, chief learning officer and director at General Dynamics Information Technology. “Machine learning helps us better understand how each individual trainee learns and performs, so that the proper skill levels are reinforced and mastered before moving on to the next block of instruction. If needed, we can anticipate additional training that might be required depending on the learner’s current performance.”
Indeed, Serfaty sees a future in which every person has a personal electronic learning record stored in the cloud, which can aid in continuing professional learning and development throughout each person’s career – even when they change jobs and employers. “In that record is the entire story of what we’ve learned, not just at school but also at work,” he said. “Imagine that we can use this electronic learning record as a GPS, not just a map, but something that tells you where you are, where you’ve got to go, and that suggests [the best ways] to get from one place to another.”
Design Interactive is already applying machine learning to training through its ScreenAdapt system, which a Window-based training application for x-ray images. ScreenAdapt presents trainees with images, then uses an infrared eye-tracking sensor to follow their eye movements as trainees study those images. By tracking users’ ability to identify areas of concern within an image – and also identify when trainees fail to scan the entire image – the system provides invaluable personalized feedback to both trainees and instructors.
The system was developed with the help of a federal small-business innovative research (SBIR) grant.
Judging the image correctly is “just a behavioral metric,” said DeVore, of Design Interactive. Adding eye-tracking takes things a step further: “It doesn’t just know whether you identified a threat or not, but also where you looked, what your scan pattern was, what you fixated on.” Errors are classified either as a failure to scan, a failure to detect or a failure to recognize.
“The scan error is simple,” DeVore explained. “It’s ‘hey, you just didn’t look in the top left quadrant for whatever reason.’ A detection error is based off the gaze indication – so you looked there, but you just didn’t detect anything. Then a recognition error would be, ‘hey, you looked there, you had a significant fixation, but you just failed to recognize that it was, in fact, contraband.’”
The ScreenAdapt then uses that feedback to select which images to present in the future, so trainees spend their time learning to overcome their weaknesses rather than practicing tasks they’re already good at. “So if you’re already really good at identifying problems in the lower right quadrant, then you’re not going to get many of those images,” DeVore said. But if you’re having a hard time with something else, “then that’s what it’s going to give you.”
The same technology is applicable in medicine – scanning medical x-rays, mammograms and other medical scans – as well as geospatial image analysis, material quality control (such as reviewing welds for cracks) and a host of other applications.
“Medicine is something we’re really interested in,” DaVore said. “We’ve spoken with a lot of radiologists and they’ve said there’s a real difference between the way a novice will scan a medical image versus an experienced performer. And there are lots of errors that occur when novices are performing versus [experienced ones]. So we definitely think our solution can accelerate the training process.”