Handling non-standard cases in computing: from user and model perspectives

Technology surrounds us, yet systems often fail when facing unexpected situations. The core idea is that well-functioning systems work in cycles, taking in information, responding, and adjusting their behavior based on the feedback they receive. But when they meet situations they were not designed for, this cycle can break down. I studied this problem from two complementary perspectives: the user’s and the model’s. From the user's perspective, I focused on people whose abilities differ from what systems commonly expect. Through user studies, I showed that offering multiple ways to interact makes technology equally accessible regardless of a person's abilities. I then built a framework that lets systems automatically detect those abilities and adapt on their own, shifting the burden of adapting away from users. This framework was validated in the context of visual impairments. I also co-developed a tool that helps people who stutter speak more fluently. Here, technology did not only adapt to users but actively supported them. From the model's perspective, I examined how machine learning systems struggle when the data they encounter does not match the data they were trained with. I introduced a new measure to accurately quantify data imbalance and its impact on performance. I also analyzed why emotion-recognition systems fail to generalize, identifying weaknesses and proposing concrete improvements. Together, these contributions offer practical steps toward technology that works for the full diversity of people and conditions, not just the ones it was originally built for.