Research Project Title:
GAN Model Inversion
abstract:The real world distribution of data is inherently long-tailed, and the data that are under-represented are often the most valuable. Yet, neural networks that are deployed in the wild are often trained on carefully curated data that minimizes the distributional biases that exist in the world. Neural networks, therefore, fail to robustly operate on observations that belong in the long-tail distribution. However, with the recent advances in neural generative models, synthesizing data has become a promising tool for generating artificial data. We propose to leverage generative adversarial networks(GANs), a successful family of generative models, as a way to generate data in parameterized fashion.
Through SuperUROP, I want to gain experience in computer vision and basically the recent advances in this particular technology and how it will affect the life of my children in the future.