Drefs, Jakob and Guiraud, Enrico and Panagiotou, Filippos and Lücke , Jörg (2023) Direct evolutionary optimization of variational autoencoders with binary latents. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2022, September 19–23, 2022, Grenoble, France.

Restricted to Repository staff only Volltext (1935Kb)

Abstract

Many types of data are generated at least partly by discrete causes. Deep generative models such as variational autoencoders (VAEs) with binary latents consequently became of interest. Because of discrete latents, standard VAE training is not possible, and the goal of previous approaches has therefore been to amend (i.e, typically anneal) discrete priors to allow for a training analogously to conventional VAEs. Here, we divert more strongly from conventional VAE optimization: We ask if the discrete nature of the latents can be fully maintained by applying a direct, discrete optimization for the encoding model. In doing so, we sidestep standard VAE mechanisms such as sampling approximation, reparameterization and amortization. Direct optimization of VAEs is enabled by a combination of evolutionary algorithms and truncated posteriors as variational distributions. Such a combination has recently been suggested, and we here for the first time investigate how it can be applied to a deep model. Concretely, we (A) tie the variational method into gradient ascent for network weights, and (B) show how the decoder is used for the optimization of variational parameters. Using image data, we observed the approach to result in much sparser codes compared to conventionally trained binary VAEs. Considering the for sparse codes prototypical application to image patches, we observed very competitive performance in tasks such as ‘zero-shot’ denoising and inpainting. The dense codes emerging from conventional VAE optimization, on the other hand, seem preferable on other data, e.g., collections of images of whole single objects (CIFAR etc.), but less preferable for image patches. More generally, the realization of a very different type of optimization for binary VAEs allows for investigating advantages and disadvantages of the training method itself. And we here observed a strong influence of the method on the learned encoding with significant impact on VAE performance for different tasks.

Item Type:	Conference or Workshop Item (Paper)
Divisions:	Faculty of Medicine and Health Sciences > Department of Medical Physics and Acoustics
Date Deposited:	01 Oct 2024 10:52
Last Modified:	01 Oct 2024 10:52
URI:	https://oops.uni-oldenburg.de/id/eprint/6352
URN:	urn:nbn:de:gbv:715-oops-64330
DOI:
Nutzungslizenz:

Actions (login required)

View Item