We introduce a concept of the generative adversarial network and consider its possibility of medical application.
We can recognize what information means by processing in our brain from input mainly through our auditory and visionary sensors.
This fact is based on long-term evolution and learning.
Sensing is attained during evolution but recognition is obtained during learning and repeating over long-term experiments.
In the machine learning, we make computers learn from various input and its relationship to output.
We input images and audios into computers such as transmission of stimulation into visionary and auditory sensors.
We make various types of neural networks for classification and regression for representing fed data adequately while resembling the learning process in human beings.
Recent advance in machine learning leads to outstanding performance beyond human beings for specific tasks.
Next-generation technique of machine learning is development of the generative network.
The generative network can yield some meaningful information from given stimulation into it.
It is very similar to imagination in human brain.
A fascinating idea proposed by a pioneering work, generative adversarial network (GAN), by Ian Goodfellow is making a race between different-role networks [1].
One is a generative network, which make better data resembling the truth from given dataset and the other is a classifier network, which strictly discriminate the generated data from truth.
Both of the networks grows up by receiving feedback through their learning and the race.
In this talk, we introduce the framework of the generative adversarial network.
In addition, we discuss the possibility of medical application of this architecture, while looking at the work on the unpaired image-to-image translation through Cycle GAN [2].
[1] I. Goodfellow et al: Advances in Neural Information Processing Systems 27 (NIPS 2014)
[2] Jun-Yan Zhu et al: 2017 IEEE International Conference on Computer Vision (ICCV), Venice, 2017, pp. 2242-2251.