Efficient prediction of trait judgments from faces using deep neural networks
Judgments of people from their faces are often invalid but influence many social decisions (e.g., legal sentencing), making them an important target for automated prediction. Direct training of deep convolutional neural networks (DCNNs) is difficult because of sparse human ratings, but features obtained from DCNNs pre-trained on other classifications (e.g., object recognition) can predict trait judgments within a given face database. However, it remains unknown if this latter approach generalizes across faces, raters, or traits. Here we directly compare three distinct types of face features, and test them across multiple out-of-sample datasets and traits. DCNNs pre-trained on face identification provided features that generalized the best, and models trained to predict a given trait also predicted several other traits. We demonstrate the flexibility, generalizability, and efficiency of using DCNN features to predict human trait judgments from faces, providing an easily scalable framework for automated prediction of human judgment.