An Analysis of Ability in Deep Neural Networks
share
Deep neural networks (DNNs) have made significant progress in a number of Machine Learning applications. However without a consistent set of evaluation tasks, interpreting performance across test datasets is impossible. In most previous work, characteristics of individual data points are not considered during evaluation, and each data point is treated equally. Using Item Response Theory (IRT) from psychometrics it is possible to model characteristics of specific data points that then inform an estimate of model ability as compared to a population of humans. We report the results of several experiments to determine how different Deep Neural Network (DNN) models perform under different training circumstances with respect to ability. As DNNs train on larger datasets, performance begins to look like human performance under the assumptions of IRT models. That is, easy questions start to have a higher probability of being answered correctly than harder questions. We also report the results of additional analyses regarding model robustness to noise and performance as a function of training set size that further inform our main conclusion
READ FULL TEXT
