Artificial intelligence-powered technologies, such as computer vision systems and voice-activated interfaces, hold considerable promise for enhancing the lives of individuals with disabilities; however, a significant issue exists in that these AI models are frequently developed utilizing limited data originating from the populations they are intended to serve.

Microsoft is collaborating with numerous nonprofit organizations to ensure these technologies accurately reflect the requirements and daily experiences of individuals living with conditions like visual impairment and limited physical mobility.

For instance, consider a computer vision system designed to identify objects and articulate their presence, such as items on a surface. It is probable that the algorithm was trained using data gathered by individuals without disabilities, representing their perspective—typically from a standing position.

An individual using a wheelchair attempting the same task may discover the system’s effectiveness is diminished due to the lower vantage point. Likewise, a person with visual impairment may struggle to position a camera correctly and maintain that position for the duration needed for the algorithm to function, necessitating a process of trial and error.

Alternatively, consider a facial recognition algorithm intended to determine a user’s attentiveness to a screen. What is the probability that the training dataset included a substantial number of individuals utilizing equipment such as ventilators, puff-and-blow controllers, or head-mounted supports that partially obscure their faces? Such “interfering factors” can substantially reduce accuracy if the system has not encountered similar scenarios.

Facial recognition systems exhibiting failures with individuals of darker skin tones, or demonstrating reduced precision with women, serve as a well-known illustration of this “garbage in, garbage out” phenomenon. A less frequently discussed, yet equally crucial, aspect is the representation of individuals with disabilities, or their unique perspectives, within these datasets.

Microsoft has announced several initiatives, jointly led with advocacy groups, aimed at addressing this “data gap” and promoting greater inclusivity in AI development.

The first is a partnership with Team Gleason, an organization dedicated to raising awareness of amyotrophic lateral sclerosis (ALS), a progressive neurodegenerative disease (the organization is named after former NFL player Steve Gleason, who received a diagnosis of the condition several years ago).

Their primary concern centers on the challenges with facial recognition described above. Individuals living with ALS experience a diverse range of symptoms and utilize various assistive technologies, which can disrupt algorithms that have not been exposed to them previously. This presents a problem if, for example, a company intends to release gaze-tracking software reliant on facial recognition, a capability Microsoft is actively pursuing.

“Computer vision and machine learning frequently do not account for the use cases and appearances of individuals with ALS and other conditions,” explained Blair Casey of Team Gleason. “Each person’s situation is unique, and their approach to technology is equally individualized. People consistently devise inventive methods to maximize efficiency and comfort.”

Project Insight is a new collaborative undertaking with Microsoft that will gather facial imagery from volunteer users with ALS as they engage in their daily activities. This data will eventually be incorporated into Microsoft’s existing suite of cognitive services and also made publicly available to enable others to refine their own algorithms.

The anticipated release date is late 2021. Microsoft’s Mary Bellard, from the company’s AI for Accessibility team, clarified that the extended timeline is due to the importance of starting from the ground up and ensuring accuracy.

“Research generates insights, which lead to models that engineers integrate into products. However, we require data to achieve the necessary level of accuracy for product integration,” she stated. “The data will be shared—this is not about enhancing a single product, but about accelerating research in these complex areas. This is work we do not wish to undertake in isolation.”

Another area for improvement lies in acquiring images from users who interact with technology in unconventional ways. Similar to the visually impaired individual or wheelchair user mentioned earlier, there is a scarcity of data representing their perspectives. Two initiatives are underway to address this need.

One involves collaboration with City University of London to expand and ultimately publicly release the Object Recognition for Blind Image Training project, which is compiling a dataset for identifying common objects—such as beverage cans and keyrings—using a smartphone camera. Distinctively, this dataset will be sourced exclusively from visually impaired users, allowing the algorithm to learn from the outset to process the type of data it will receive in real-world applications.

The other is an expansion of VizWiz to better incorporate this type of data. This tool assists individuals who require immediate help with tasks such as determining the expiration date of a yogurt container or verifying the presence of a vehicle in the driveway. Microsoft partnered with the app’s creator, Danna Gurari, to enhance the app’s existing database of tens of thousands of images, along with their corresponding questions and captions. They are also developing features to notify users when their images are too dark or blurry for analysis or submission.

Achieving inclusivity is a complex undertaking, as it involves people and systems that may unintentionally define “normal” and subsequently fail to function effectively outside those parameters. For AI to be truly inclusive, the definition of “normal” must be broadened, which will require substantial effort. Until recently, this issue was largely unaddressed, but that is now changing.

“This is something the ALS community has needed for years,” said Casey. “This technology already exists—it’s simply not being utilized. By discussing it, we can encourage further development, which is essential for the community as a whole.”