Reproduce “Landscape Learning for Neural Network Inversion” on OpenResearch

About this paper

Ruoshi Liu, Chengzhi Mao, Purva Tendulkar, Hao Wang, Carl Vondrick

Researchers from Columbia University and Rutgers University introduced a framework that learns a remapped latent space for Optimization-Based Inference (OBI), enabling faster and more stable neural network inversion. The system improves OBI efficiency by up to 10 times in GAN inversion, 19% in 3D human pose reconstruction, and 18% in robust accuracy for adversarial defense, by creating smoother loss landscapes.

Key insights

• Learning a remapping of the input space can significantly smooth and steepen the loss landscape, enabling faster and more stable gradient descent.
• A coordinate descent training algorithm with an experience replay buffer effectively optimizes the mapping network by learning from accumulated optimization trajectories.
• Even a randomly initialized mapping network can offer some performance improvements, suggesting that merely introducing a new latent space with specific properties can be beneficial.

Summary by alphaXiv.