ELBD: Efficient score algorithm for feature selection on latent variables of VAE
This addresses feature selection in variational autoencoders for researchers working with generative models, though it appears incremental as an extension of existing VAE methodology.
The authors developed ELBD (evidence lower bound difference) for feature selection on VAE latent variables, proposing efficient algorithms that improve model performance by weighting important latent variables. Experimental results on 7 generative datasets showed effectiveness compared to 9 other methods, with extensions to 5 classification datasets also yielding satisfactory results.
In this paper, we develop the notion of evidence lower bound difference (ELBD), based on which an efficient score algorithm is presented to implement feature selection on latent variables of VAE and its variants. Further, we propose weak convergence approximation algorithms to optimize VAE related models through weighing the ``more important" latent variables selected and accordingly increasing evidence lower bound. We discuss two kinds of different Gaussian posteriors, mean-filed and full-covariance, for latent variables, and make corresponding theoretical analyses to support the effectiveness of algorithms. A great deal of comparative experiments are carried out between our algorithms and other 9 feature selection methods on 7 public datasets to address generative tasks. The results provide the experimental evidence of effectiveness of our algorithms. Finally, we extend ELBD to its generalized version, and apply the latter to tackling classification tasks of 5 new public datasets with satisfactory experimental results.