Diversity of thought is very important in industrial design. If no one thinks about designing technology for multiple body types, people can get hurt. The invention of the seat belt is an often cited example of this phenomenon. Seat belts are designed based on crash dummies, which traditionally have male proportions and reflect the bodies of the team members they work with.
The same phenomenon is currently occurring in the field of motion capture technology. Throughout history, scientists have sought to understand how the human body moves. But how do we define the human body? Decades ago, many studies evaluated “healthy male” subjects. Others used surprising models like dismembered corpses. Even today, some of the latest research used to design fall detection technology relies on methods such as hiring stunt actors to fake falls.
Over time, various flawed assumptions were codified into the standards for motion capture data used to design some AI-based technologies. New research recently published as a preprint and scheduled to be presented at the Human Factors in Computing conference shows that these flaws make AI-based applications less difficult for people who don't fit into preconceived “typical” body types. This means it may not be safe. May system.
“We have thoroughly investigated the so-called gold standards used in all kinds of research and design, many of which are flawed or focused on very specific body types. ,” said study co-author and assistant professor at the University of Michigan's School of Information and Center for Complex Systems Research. “We want engineers to be aware of how these social aspects are encoded into technical aspects, hidden in mathematical models that appear objective or infrastructural. I want to.”
Jacobs said this is a critical time for AI-based systems, and there may still be time to spot and avoid potentially dangerous assumptions being codified into AI-informed applications. It states that there is no.
Motion capture systems create representations of objects by collecting data from sensors placed on the objects and recording how these objects move through space. These schematics become part of the tools used by researchers, including open-source libraries of movement data and measurement systems aimed at providing baseline standards for human body movement. Developers are increasingly using these baselines to build all kinds of AI-based applications. These include fall detection algorithms for smartwatches and other wearables, self-driving cars that need to detect pedestrians, computer-generated images for movies and video games, and manufacturing equipment that interacts safely. such as human workers.
“Many researchers don't have access to advanced motion capture labs to collect data, so they've increasingly relied on benchmarks and standards to build new technology,” Jacobs says. “However, if these benchmarks do not include representation of all bodies, especially those who may be involved in real-world use cases, such as older adults who may be at risk for falls, then these standards have significant There may be a defect.”
She talks about what we can learn from past mistakes, such as cameras that failed to accurately capture all skin tones, and seatbelts and airbags that failed to protect people of all shapes and sizes in car crashes. I hope.
corpse in the machine
Jacobs and colleagues from Cornell University, Intel, and the University of Virginia conducted a systematic literature review of 278 motion capture-related studies. In most cases, the researchers concluded, motion capture systems captured the movements of “people who are male, white, 'able-bodied,' and of moderate weight.”
And sometimes the bodies of these white men were dead. In reviewing research spanning three historical eras of motion capture science, dating back to the 1930s, the researchers found that motion capture influenced how scientists at the time understood the movement of body parts. I researched the project. For example, a seminal 1955 study funded by the Air Force used overwhelmingly white, male, and slender or athletic bodies to create an optimal cockpit based on the pilot's range of motion. The study also collected data from eight dismembered cadavers.
A full 20 years later, a study prepared for the National Highway Traffic Safety Administration used a similar method. Six dismembered male cadavers were used to design a vehicle impact protection system.
In most of the 278 studies examined, motion-capture systems captured the movements of “male, white, 'able-bodied', moderately heavy people.”
Although these studies are decades old, these assumptions have become entrenched over time. Jacobs and her colleagues found numerous examples of these outdated reasoning being carried over into later research and ultimately still influencing modern motion capture research.
“If you look at the technical documentation for modern systems in production, they describe the 'traditional baseline standards' that are used,” says Jacobs. “By digging into it, you start jumping through time right away. OK, this is based on this previous study, this is based on this study, this is based on this study, and ultimately Go back to Air Force research designing cockpits.'' Frozen corpses. ”
The elements underpinning technical best practices are “artificial, deliberately emphasizing the human rather than the human, and often retain biases and inaccuracies from the past,” she said. Kasia Chmielinski, Project Leader for the Nutrition Project and Digital Civil Society Fellow at Stanford University. Lab “Historical errors thus often inform the “neutral” foundations of current technological systems. This may result in software or hardware not working equally well for all populations, experiences, and purposes. ”
These issues can hinder engineers who want to get things right, Chmielinski says. “Many of these issues are built into the fundamental elements of the system, so teams innovating today may not be able to address bias and errors as quickly as they want to. ” she says. “If you're building an application that uses third-party sensors, and the sensors themselves have biases in what they detect and what they don't, what's the right thing to do?”
Jacobs says engineers need to examine their sources of “ground truth” to ensure that the gold standard they measure against is actually gold. Engineers need to consider this social evaluation as part of their job in order to design technology for everyone.
“Saying, 'We know that human assumptions are built in and can be hidden or obscured,' can affect how we choose what is included in a dataset and how we apply that in our work.” ,” says Jacobs. “This is a socio-technical thing, and engineers need a lens through which they can say, 'My system does what I say it does, and it doesn't cause unwarranted harm.'”
From an article on your site
Related articles on the web