Constrain-KL for Variational Autoencoders
A constrained optimisation approach for setting exact KL targets in beta-VAE training.
At Imperial College London, I developed Constrain-KL, a constrained optimisation method for beta-VAE training.
The method enforces an exact KL target, removing the need to tune beta. It matched NVAE accuracy on CIFAR-10 in our experiments.
