| Abstract |
I will be reporting on work in progress at CWI, where we study crowdsourcing tasks that are too difficult for the crowd. As an example, I will use the task of classifying Taiwanese coral fish species from low quality, high quantity ocean camera footage. This is a task that cannot be done by experts (too few experts, too much data), but also not by non-experts (too difficult, too domain-specific). We suspect it can be done by supervised machine learning, but not with the small quantities of training data the experts can provide. We show that we can transform the original task into a simpler, gamefied task that can be carried out by non-experts, and show how we combine the input of the experts, encyclopaedic knowledge and imperfect computations to realise this transformation, and how our approach could be applied to other domains. |