NeurIPS 2020

Network-to-Network Translation with Conditional Invertible Neural Networks

Meta Review

All the reviewers agree the work is original and potentially impactful. Two reviewers rate the paper top 15%, one rates top 50%, and one rates a good submission. The work is timely as it provides an efficient way to leverage large-scale networks trained by resourceful institutions for various novel translation applications such as text-to-image or image-to-image. If we can link the latent space learned by a GPT3 and a BigGAN, we could achieve high-quality text to image or image to text translation tasks. The major contribution is a general framework for learning the mapping between two independently learned semantic space using a conditional invertible neural network. After consolidating the reviews and rebuttal, the AC agress with the assessment and congratulate the authors for the acceptance.