A Collaborative Training Using Crowdsourcing and Neural Networks on Small and Difficult Image Classification Datasets

Although machine learning has recently achieved performance that exceeds human capacity in prediction, humans still have an advantage in difficult tasks when the number of training samples is small or when human knowledge is required to identify features that are included in samples. However, if specialized knowledge is required, the number of humans that can perform those tasks is limited. In this study, we effectively use crowdsourcing to incorporate domain knowledge in neural network trainings; specifically, we decide feature values by asking crowdsourcing workers to answer easy questions prepared based on dictionary. We evaluated this method on a single type of task that is intuitive and relevant for non-specialists, which is binary classification of dog image datasets with similar breeds, and found that using crowdsourcing tended to improve the performance of machine-learning models. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd 2022.