Understanding Deep Learning Performance through an Examination of Test Set Difficulty: A Psychometric Case Study

Deep neural networks (DNNs) have set state of the art results in many machine learning and NLP tasks. However, we do not have a strong understanding of what DNN models learn. In this paper, we examine learning in DNNs through analysis of their outputs. We compare DNN performance directly to a human population, and use characteristics of individual data points such as difficulty to see how well models perform on easy and hard examples. We investigate how training size and the incorporation of noise affect a DNN's ability to generalize and learn. Our experiments show that unlike traditional machine learning models (e.g., Naive Bayes, Decision Trees), DNNs exhibit human-like learning properties. As they are trained with more data, they are more able to distinguish between easy and difficult items, and performance on easy items improves at a higher rate than difficult items. We find that different DNN models exhibit different strengths in learning and are robust to noise in training data.