Teachable machines for accessibility

Abstract

How can accessibility research leverage advances in machine learning and artificial intelligence with limited data? In this article, we argue that teachable machines can empower accessibility research by enabling individuals with disabilities to personalize a data-driven assistive technology. By significantly constraining the conditions of the machine learning task to a specific user and their environment, these technologies can achieve higher robustness in real world scenarios. In contrast to automatic personalization, the end user is called to consciously provide training examples and actively interact with the machine learning algorithm to increase its accuracy. We demonstrate this concept with a concrete example: teachable object recognizers trained by and for blind users. Furthermore, we discuss open challenges in designing and building teachable machines with a focus on accessibility.