Classful Object Generation Using Transposed Convolutional Network Fine-Tuned with a Classifier

On-demand generation of object that belongs to a specific class is of great importance in various AI applications. We show it is possible to generate on demand a object that belongs to a particular class from a single low-volume vector by using Tansposed Convolutional Network (TCN). TCN is an up-sampling method that can learn from experiences to optimally adjust its parameters to retain relevant information contained in those experiences. The network is first pre-trained using supervised learning to set its parameters in a good region and then fine-tuned with a trained classifier. The low volume input vector to the network is made up of two fixed-sized segments. The first segment employs one-hot scheme to indicate class identity of the object, while the second segment is composed of elements with their values randomly selected. The purpose of this second segment is to inject randomness in the object generation process and hopefully to capture some interesting features (such as style) in the training data. The contribution of this work is two-fold: 1) show that complexed, high dimensional, classiful object can be learned and generated on demand from a short low dimensional vector; 2) propose a fine tuning procedure to fine-tune the generator with a trained classifier to improve the quality of object generated. As a proof of concept, a generator network is built and trained to produce images of any hand-written digit on demand.