Kaitav Mehta

 

Date:  Monday April 8th, 2019

Time:  11: 00 am – 1:00 pm

Location: Essex Hall Room 122

 

Recent developments in Deep Learning field is noteworthy when it comes to learning probability distribution of points through neural networks, and one of key part for such progress is because of Generative Adversarial Networks (GANs). In which two neural networks (Generator and Discriminator) compete among each other to learn probability distribution of points in visual pictures. Lots of research has been done to overcome the shortage of GANs like training instability, mode collapse and convergence of neural networks. But there was no significant proof found if modern techniques consistently outperform vanilla GANs, and it turns out that different modern techniques perform differently on different datasets. In this paper, we propose an evolutionary training technique using cultural algorithm (CA) for neuro-evolution of deep task oriented GANs architecture to find the best architecture for the given dataset. Using fitness function, we select the best architecture generated by the evolutionary algorithm in a fixed number of generations and train best architecture evolved from CA for a higher number of epochs. We have compared our approach with the Genetic Algorithm (GA) based neuro-evolution of GANs and we show that CA based neuro-evolution of GANs evolves architecture which can generate higher number of stroke-face images with better resolution when there is fewer data of original stroke faces.

 

Thesis Committee: