Simultaneous generation-classification using LSTM
Daniel Marino, Kasun Amarasinghe, Milos Manic · 2016
The idea that a concept is properly learned by an agent when the agent is able to generate examples and non-examples of the concept, has motivated research on generative models. Generative models are trained with the aim of improving performance of tasks such as classification. In this paper, a Long Short Term Memory (LSTM) architecture for simultaneous generation-classification is presented. The architecture is designed with the purpose of serving as a model which can generate sequence samples, while simultaneously classifying a given sequence. The presented generation-classification methodology was implemented on a sentiment analysis task. However, it can be applied to any sequence modelling or classification task. The experimental results suggest that this approach can be particularly useful as a regularization methodology which acts similarly to pre-training through Restricted Boltzmann Machines or auto-encoders.