Balancing reconstruction error and Kullback-Leibler divergence in Variational Autoencoders

In the loss function of Variational Autoencoders there is a well known\ntension between two components: the reconstruction loss, improving the quality\nof the resulting images, and the Kullback-Leibler divergence, acting as a\nregularizer of the latent space. Correctly balancing these two components is a\ndelicate issue, easily resulting in poor generative behaviours. In a recent\nwork, Dai and Wipf obtained a sensible improvement by allowing the network to\nlearn the balancing factor during training, according to a suitable loss\nfunction. In this article, we show that learning can be replaced by a simple\ndeterministic computation, helping to understand the underlying mechanism, and\nresulting in a faster and more accurate behaviour. On typical datasets such as\nCifar and Celeba, our technique sensibly outperforms all previous VAE\narchitectures.\n