Random Inception Module and Its Parallel Implementation

Abstract

Inception module is proposed in GoogLeNet, which improves performance by increasing the width of the network. Multiple branches are computed in parallel, which makes the inception module naturally take the advantage of GPU high-performance computing. In this paper, we propose a parallel implementation of the inception module to accelerate the training and test of the inception networks. However, convolution neural networks are prone to overfitting due to the huge amount of parameters. We propose random inception module to avoid overfitting and accelerate inception module. In order to demonstrate the effectiveness of the proposed methods, we compare the performance of the random inception module with original inception module on CIFAR-10 dataset. Experimental results show our parallel inception module obtains over 2.8\(\times \) speedup compared with Caffe. And our proposed RIM indeed behaves in a manner of regularization and speeds up convergence.