Nov 04, 2016 (modified: Dec 10, 2016)ICLR 2017 conference submissionreaders: everyone
Abstract:Convolutional Neural Networks have proved to be very efficient in image and audio processing. Their success is mostly attributed to the convolutions which utilize the geometric properties of a low - dimensional grid structure. This paper suggests a generalization of CNNs to graph-structured data with varying graph structure, that can be applied to standard regression or classification problems by learning the graph structure of the data. We propose a novel convolution framework approach on graphs which utilizes a random walk to select relevant nodes. The convolution shares weights on all features, providing the desired parameter efficiency. Furthermore, the additional computations in the training process are only executed once in the pre-processing step. We empirically demonstrate the performance of the proposed CNN on MNIST data set, and challenge the state-of-the-art on Merck molecular activity data set.
TL;DR:A generalization of CNNs to standard regression and classification problems by using random walk on the data graph structure.
Keywords:Supervised Learning, Deep learning
Enter your feedback below and we'll get back to you as soon as possible.