CNTK-based multi-task learning for gender and age recognition using face images

Abstract

Since the success of deep learning in numerous contests in pattern recognition and machine learning, it has already been used in widespread fields such as speech recognition, medical diagnosis and unmanned vehicle. The improvement of computing facilities, big data and sophisticated deep learning techniques have led to the success of deep learning. In addition, the deep learning software such as Caffe, Theano, Torch and etc. also lead to the popularization of deep learning. In this paper, we explored the CNTK (computational network toolkit or Microsoft cognitive toolkit), that is a deep learning software provided by Microsoft. The CNTK is a general purpose machine learning software that can support multi-devices and multi-GPUs. That can model various neural network architectures such as deep neural network, convolutional neural network and recurrent neural network. Also, researcher can easily design and evaluate various neural network architectures using BrainScript, C++ and Python. We used the CNTK for gender and age recognition by face images with multi-task learning. Face images contain rich information such as identity, expression, age, gender and races. It is helpful to extract multiple information simultaneously from face images for developing applications. In addition, it is observed that multi-task learning may help to improve the performance than single-task learning. We modeled the simple multilayer perceptron network for gender and age recognition. And then we applied multi-task learning with a same network. For the computational evaluation, we collected a set of face images from FERET (the facial recognition technology) database for gender and age recognition. In the computational experiments, we confirmed that the performance of multi-task learning is slightly better than single-task learning. We also confirmed that it is easy to model the neural network architecture using the CNTK.