Abstract
A large number of works aim to alleviate the impact of noise due to anunderlying conventional assumption of the negative role of noise. However, someexisting works show that the assumption does not always hold. In this paper, weinvestigate how to benefit the classical models by random noise under theframework of Positive-incentive Noise (Pi-Noise). Since the ideal objective ofPi-Noise is intractable, we propose to optimize its variational bound instead,namely variational Pi-Noise (VPN). With the variational inference, a VPNgenerator implemented by neural networks is designed for enhancing base modelsand simplifying the inference of base models, without changing the architectureof base models. Benefiting from the independent design of base models and VPNgenerators, the VPN generator can work with most existing models. From theexperiments, it is shown that the proposed VPN generator can improve the basemodels. It is appealing that the trained variational VPN generator prefers toblur the irrelevant ingredients in complicated images, which meets ourexpectations.