Convolutional Neural Networks for Age and Gender Classification

This paper focuses on the problem of gender and age classification for an image. I build off of previous work [12] that has developed efficient, accurate architectures for these tasks and aim to extend their approaches in order to improve results. The first main area of experimentation in this project is modifying some previously published, effective architectures used for gender and age classification [12]. My attempts include reducing the number of parameters (in the style of [19]), increasing the depth of the network, and modifying the level of dropout used. These modifications actually ended up causing system performance to decrease (or at best, stay the same) as compared with the simpler architecture I began with. This verified suspicions I had that the tasks of age and gender classification are more prone to over-fitting than other types of classification. The next facet of my project focuses on coupling the architectures for age and gender recognition to take advantage of the gender-specific age characteristics and agespecific gender characteristics inherent to images. This stemmed from the observation that gender classification is an inherently easier task than age classification, due to both the fewer number of potential classes and the more prominent intra-gender facial variations. By training different age classifiers for each gender I found that I could improve the performance of age classification, although gender classification did not see any significant gains.