Multitask variational autoencoders

Abstract

Variational autoencoders (VAEs) are widely used for generative modeling and repre- sentation learning tasks. This thesis presents two novel approaches aimed at enhancing the performance of VAEs through the integration of semi-conditional variational autoen- coders (SCVAEs). The integration of SCVAEs and VAEs is motivated by the potential for improvement in the effectiveness of capturing the underlying data distribution and the improvement in generating high-quality samples. The first method extends the traditional VAEs by incorporating a second conditioned decoder, thereby enabling the model to multi-task and learn better latent representations. The second method utilizes a unified decoder for both tasks by employing sophisticated training strategies. These approaches are implemented and evaluated on Gaussian VAEs and VQ-VAEs. Extensive experimentation is conducted across diverse image datasets, including MNIST, CIFAR10, and CelebA. The results show that, in certain cases, the proposed methods yield superior performance compared to standard VAE architectures. By bridg- ing the gap between SCVAEs and VAEs, this work gives new insights on how to improve the methods further and opens up new avenues for future research in the field of generative modeling.