Handwritten Digit Recognition using Sparse Deep Architectures

A lot of research has been lately focusing on deep neural networks as an alternative to shallow ones. The added advantage among many, is the automated feature extraction of pattern from data. These models have been applied successfully to many tasks, including handwritten digit recognition, where they lead the state of the art performance. In this paper we apply a sparse deep belief network and a denoising autoencoder to a new dataset proposed in the ICDAR 2013 handwritten digit competition, which is a challenging alternative in multiple aspects to many popular digit datasets. Additionally we raise some difficulties met during modelling. We finally put the spot on some elements that could improve the performance for subsequent attempts in improving models' accuracy.