OPTIMIZATION OF DISTRIBUTED CONVOLUTIONAL NEURAL NETWORK FOR IMAGE LABELING ON ASYNCHRONOUS GPU MODEL

The problem of image labeling consists in assigning semantic labels to every pixel in an image. Recently, given their capacity to learn rich features, many state-of-the-art techniques make use of Convolutional Neural Networks (CNNs) for dense image labeling tasks (e.g., multi-class semantic segmentation). In this paper, we propose the optimization of distributed CNN for image labeling on asynchronous GPU (Graphics Processing Unit) model. We deduce equations for ADMM (Alternating Direction Method of Multipliers) which has been widely used in variety of optimization problems, and apply the optimization to distributed convolutional neural networks training. Our proposed framework is designed and implemented in GPU-GPU communication depending on the asynchronous model. The results show that we could get good pixel accuracy of image labeling while having a fast labeling speed.