Training Deep Neural Nets to Aggregate Crowdsourced Responses

We propose a new method for aggregating crowdsourced responses, based on a deep neural network. Once trained, the aggregator network gets as input the responses of multiple participants to the same set of questions, and outputs its prediction for the correct response to each question. We empirically evaluate our approach on a dataset of responses to a standard IQ questionnaire, and show it outperforms existing state-of-the-art methods.