- Abstract: Recent advances in Generative Adversarial Networks facilitated by improvements to the framework and successful application to various problems has resulted in extensions to multiple domains. IRGAN attempts to leverage the framework for Information-Retrieval (IR), a task that can be described as modeling the correct conditional probability distribution p(d|q) over the documents (d), given the query (q). The work that proposes IRGAN claims that optimizing their minimax loss function will result in a generator which can learn the distribution, but their setup and baseline term steer the model away from an exact adversarial formulation, and this work attempts to point out certain inaccuracies in their formulation. Analyzing their loss curves gives insight into possible mistakes in the loss functions and better performance can be obtained by using the co-training like setup we propose, where two models are trained in a co-operative rather than an adversarial fashion.
- Keywords: GAN, Deep Learning, Reinforcement Learning
- TL;DR: Points out problems in loss function used in IRGAN, a recently proposed GAN framework for Information Retrieval. Further, a model motivated by co-training is proposed, which achieves better performance.