Transfer learning for extracting age-related features from face images

Recent advances in the software and hardware supporting Deep Neural Networks have led to increases in the accessibility and applicability of the technology. A key area of growth is the use of neural networks to predict health-related endpoints such as mortality, morbidity, and other more abstract biological endpoints.

In this talk, I will describe the process of extracting age-related health information from facial imagery and linking it to the accepted health endpoints. I will also discuss the bias associated with improperly replicated face alignment (which degrades prediction accuracy) and the models resilience to such misalignments. Finally, I will present initial results showing the transferability of deep face recognition features to prediction of dermatological endpoints, and discuss the problem of model overloading with higher resolution images at inference time.