Pixel-Based LSTM Generative Model

Abstract

Applying computational intelligence techniques to create generative models of digits or alphabets has received somewhat little attention as compared to classification task. It is also more challenging to create a generative model that could successfully capture styles and detailed characteristics of symbols. In this paper, we describe the application of the Long Short-Term Memory (LSTM) model trained using a supervised learning approach for generating a variety of the letter A. LSTM is a recurrent neural network with a strong salient feature in its ability to handle long range dependencies, hence, it is a popular choice for building intelligent applications for speech recognition, conversation agent and other problems in time series domains. To formulate the problem as a generative task, all the pixels in a 2D image representing an alphabet (i.e., the letter A in this study) are flattened into a long vector to train the LSTM model. We have shown that LSTM has successfully learned to generate new letters A showing many coherent stylistic features with the original letters from the training sets.