Deep Learning in Character Recognition Considering Pattern Invariance Constraints

Character recognition is a field of machine learning that has been under research for several decades. The particular success of neural networks in pattern recognition and therefore character recognition is laudable. Research has also long shown that a single hidden layer network has the capability to approximate any function; while, the problems associated with training deep networks therefore led to little attention given to it. Recently, the breakthrough in training deep networks through various pre-training schemes have led to the resurgence and massive interest in them, significantly outperforming shallow networks in several pattern recognition contests; moreover the more elaborate distributed representation of knowledge present in the different hidden layers concords with findings on the biological visual cortex. This research work reviews some of the most successful pre-training approaches to initializing deep networks such as stacked auto encoders, and deep belief networks based on achieved error rates. More importantly, this research also parallels investigating the performance of deep networks on some common problems associated with pattern recognition systems such as translational invariance, rotational invariance, scale mismatch, and noise. To achieve this, Yoruba vowel characters databases have been used in this research.