Understanding Human Aging Patterns from a Machine Perspective

Abstract

Recent research shows that the aging patterns deeply learned from large-scale data lead to significant performance improvement on age estimation. However, the insight about why and how deep learning models achieved superior performance is inadequate. In this paper, we propose to analyze, visualize and understand the deep aging patterns. We first train a series of convolutional neural networks for age estimation, and then illustrate the learning outcomes using feature maps, activation histograms, and deconvolution. We also develop a visualization method that can compare the facial appearance and track its changes at different ages through the mapping between 2D images and a 3D face template. Our framework provides an innovative way to understand human facial aging process from a machine perspective.