Researchers are in the process of building diverse training data sets that include information from disabled people such as people with low vision, who use a wheelchair or have conditions like ALS.
Disabled people are often using technology early on in order to accomplish everyday tasks. Artificial Intelligence (AI) drives many of these services like voice and object recognition. However, in many cases, these products are trained on data from able-bodied people, meaning that these algorithms have a very limited understanding of body types, communication styles, and facial expressions.
Microsoft researchers are working with advocacy groups in order to solve this problem and build data that can better reflect all types of users and real-life scenarios. AI models need to be adapted to diverse situations and be able to correctly identify the needs of disabled people. Whether it is reading cues from people with autism or handwriting from people with Parkinson’s disease or tremors, these algorithms have to be wired for all different cases. This brings out the question of how ‘normal’ is defined by AI systems and who is in charge of that.
Moreover, Microsoft researchers are also studying how often public datasets used to train AI systems to include data from people over 80 years old. Having data from disabled and older people will make algorithms smarter and more adequate.
Microsoft is thus funding a few projects in order to make AI more suitable for everyone, including:
- Object Recognition for Blind Image Training (ORBIT): this aims to build a public data set from images taken by people who are blind or have low vision so it can personalize image recognition and identify particular objects.
- VizWiz data set: the goal of this is to work with people who are blind or have low vision to better understand their expectations of AI captioning tools and improve how computer vision algorithms interpret photos taken by them.
- Project Insight: this aims to create an open dataset of facial imagery of people living with ALS in order to enhance computer vision and train AI models on a broader dataset.