Advanced deep neural networks for pattern recognition: An experimental study

Abstract

Difficulty in pattern recognition is perceptible and neural networks approach the problem by way of learning from similar known patterns. Interest in Neural Networks started in the early 1980s when they were deemed to effectively model the human thought process. Speech recognition which first used Artificial Neural Networks (ANNs) to model the states of a Hidden Markov Models (HMMs) later started using Gaussian Mixture Models (GMMs). GMM-HMM systems have been the standard until recently when a new concept of Deep Neural Networks (DNNs) pre-trained using Restricted Boltzmann Machines (RBMs) came into existence. The discriminative capability of the resulting DNN is found to improve the performance of the recognition systems. The experimental work with DNN for recognizing patterns in handwriting and speech corpus has been carried out. In this work we implemented Deep Neural Networks for the above tasks and the pre trained DNN has been used for extracting bottleneck features and hereby improving the performance of the baseline systems with respect to recognition errors.