Pioneer Centre for AI Talk: Elijah Cole

We need computer vision systems that empower human experts working on important real-world problems. An ecologist monitoring biodiversity in Kenya and a doctor screening a population for early signs of Alzheimer’s disease face the same obstacle: they must manually inspect a vast amount of raw data to solve their problems. Computer vision promises a solution: if the expert annotates enough data, we can train an algorithm to automatically extract important information from images. While the computer vision community has made impressive progress over the last few years, most of this progress has been measured against generic benchmark datasets like ImageNet. This progress may not translate to expert tasks. For instance, we have found that recent self-supervised learning methods still need work to solve fine-grained classification tasks. Throughout this talk, I will show that expert tasks present exciting opportunities (e.g. useful domain knowledge is available) and challenges (e.g. high-quality labels are scarce, fine-grained categories) that generic tasks do not. I hope to convince you that there is a symbiotic relationship between computer vision and expert tasks. By developing new methodology, we can enable and empower experts. By studying expert tasks, we can make fundamental progress in computer vision. I will discuss some examples from my recent work related to leveraging domain knowledge and learning from limited and imperfect data. Finally, I will show how we can channel the efforts of the computer vision community towards expert tasks.