A01Multi-functional CNN by Piling up Single Functions for Emergence of New Functions
In this research, we try to confirm that CNNs have an ability to emergence new functions like human brain by training multiple functions on the same CNN and combining/piling-up them in the deploying time.
In our previous work on image style transformation, we have confirmed that the proposed conditional conv-deconv network was able to mix multiple styles, although the network was trained with each of the training styles independently. This comes from the linearity of neural networks. In addition, the proposed network was also able to transfer different styles to the different parts of a given image at the same time, that is “spatial style transfer”. This comes from the locality of CNNs. From these observations, we think we can combine multiple different functions if we can train different functions with the same CNN, and we expect interesting and useful functions are emerged by combining various function. This will bring one step for realization of artificial general intelligence.