AI summary

This paper introduces Multi-Dimensional Recurrent Neural Networks (MDRNNs) to extend RNNs to multi-dimensional data, addressing scaling issues and improving robustness to input warping for tasks like image segmentation, demonstrating effectiveness on Air Freight and MNIST datasets.

Abstract

Recurrent neural networks (RNNs) have proved effective at one di- mensional sequence learning tasks, such as speech and online handwriting recog- nition. Some of the properties that make RNNs suitable for such tasks, for exam- ple robustness to input warping, and the ability to access contextual information, are also desirable in multi-dimensional domains. However, there has so far been no direct way of applying RNNs to data with more than one spatio-temporal dimension. This paper introduces multi-dimensional recurrent neural networks, thereby extending the potential applicability of RNNs to vision, video process- ing, medical imaging and many other areas, while avoiding the scaling problems that have plagued other multi-dimensional models. Experimental results are pro- vided for two image segmentation tasks.