Abstract by Joe Davison
Inferring Individual Features from Face Identity Embeddings
Advances in techniques for training Convolutional Neural Networks have enabled significant progress in a broad array of image recognition problems, including face recognition. The standard approach to discerning facial similarity involves using neural networks which embed images of individuals into a hyperdimensional Euclidian space which contains the fundamental identifying information for an individual. In my research, I explore the interpretability of these embeddings and their use in inferring other information about the individual. So far, I show success in predicting features including race, gender, and even country of origin from these vector embeddings. This shows the usefulness of using face embedding models as a technique for highly efficient transformation of pixel information into a rich, compact feature space which can be used by a broad class of algorithms for various applications.