Reproduce “The Hessian Penalty: A Weak Prior for Unsupervised Disentanglement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
William Peebles, John Peebles, Jun-Yan Zhu, Alexei Efros, Antonio Torralba
Researchers from UC Berkeley, Yale, Adobe Research, and MIT CSAIL propose the Hessian Penalty, a regularization term that minimizes off-diagonal elements of the generator's Hessian matrix to encourage unsupervised disentanglement in deep generative models. This method generates more interpretable and controllable latent spaces, achieving lower Perceptual Path Length (PPL) on various datasets and enabling the discovery of disentangled directions in pre-trained models.
Key insights
- Encouraging a diagonal Hessian matrix acts as a 'weak prior' that effectively biases deep generative models towards learning disentangled and interpretable representations.
- The Hessian Penalty's efficient stochastic approximation allows its application to virtually any deep generative model without architectural modifications or auxiliary encoders.
- The penalty promotes 'latent space shrinkage' in overparameterized models, leading to sparsity where redundant latent components become inactive.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.