Core Concepts
Any dimension-preserving neural network can be trained as a generative model through maximum likelihood estimation by using an efficient gradient estimator for the change of variables formula.
Abstract
The paper introduces free-form flows (FFF), a new approach to training normalizing flow models that removes the architectural constraints typically required for analytical invertibility and tractable Jacobian computations.
The key innovation is an efficient gradient estimator that allows training any dimension-preserving neural network as a generative model through maximum likelihood optimization. This is achieved by learning an approximate inverse of the encoder network, and using a reconstruction loss to ensure the encoder-decoder pair are close to being inverses.
The authors show theoretically that optimizing this relaxed objective has the same critical points as the original maximum likelihood objective, provided the reconstruction loss is minimized. They also prove that the solutions learned by optimizing the relaxed objective exactly match the data distribution.
Experimentally, the authors demonstrate the versatility of free-form flows. On a simulation-based inference benchmark, FFF models achieve competitive performance with minimal tuning. On molecule generation tasks, the authors leverage the flexibility to use equivariant graph neural networks, outperforming previous normalizing flow approaches in terms of sample quality and generation speed.
Stats
Normalizing flows directly maximize the data likelihood by learning an invertible mapping from data to a simple latent distribution.
Traditionally, the design of normalizing flows was constrained by the need for analytical invertibility and tractable Jacobian computations.
The free-form flow approach removes these architectural constraints, allowing any dimension-preserving neural network to be trained as a generative model.
Quotes
"We overcome this constraint by a training procedure that uses an efficient estimator for the gradient of the change of variables formula."
"Our approach allows placing the emphasis on tailoring inductive biases precisely to the task at hand."