Multi-Stage Bayesian Optimization for Adversarial Fine Tuning

At the moment there are many works connected with Generative Adversarial Networks (GAN). Usually a Generative Adversarial Network is considered to consist of two neural networks: a generator and a discriminator. These networks are trained simultaneously. The generator's goal is to learn how to generate objects so that the discriminator could perceive them as real ones, and the discriminator's goal is to learn how to distinguish real objects from the objects created by the generator. Then the trained generator can be used to create new objects. This problem is well studied in its original statement. But we suppose that the generator is not a network and, moreover, there is no explicit relationship between the generator’s parameters and the objects. The example of this task may be the refinement of some physical constants according to real experiments and other inverse problems. Existing algorithms use different optimization methods to find optimal parameters, but they require a large number of generated examples. The main goal of this paper is to develop a new method of gradientless optimization, which requires fewer examples, by introducing a multi-step optimization process. So, the basic idea is to use a gradual transition from simple classifiers to complex ones, applying complex classifiers only in those areas of parameters where simple ones could not cope with the task. In this research a multi-step algorithm was created as well as experiments based on demonstration tasks were conducted. As a result, the advantage of the constructed algorithm over existing approaches was revealed. Due to the general statement, the algorithm can also be applied to other optimization problems.