کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / چهاردهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Improving Training Stability in Variational Autoencoders Through the Integration of Score Matching Loss

نویسندگان :

Amirreza Mokhtari Rad¹ Pouya Ardehkhani² Hormehr Alborzi³

1- پردیس فارابی دانشگاه تهران 2- پردیس فارابی دانشگاه تهران 3- پردیس فارابی دانشگاه تهران

کلمات کلیدی :

Variational Auto Encoder،Training،Stability،Generative Models،Score Matching

چکیده :

In this research, a Variational Autoencoder (VAE) model was developed, and the CIFAR100 dataset was employed as the primary data source. The problem addressed pertained to the instability in the training process of VAE models. To mitigate this issue, various loss expressions were explored, including the use of score matching loss independently, in conjunction with total variation loss, and in combination with reconstruction loss. The innovative approach revealed that when score matching loss was integrated either with total variation loss or when applied as a standalone loss function, the training process exhibited increased stability. This was evident through smoother loss curves and latent space visualizations that displayed characteristics akin to a normal distribution. As a consequence, this novel approach promises the potential for building more stable generative models, which can significantly enhance the overall training process in VAEs. This innovation provides a valuable contribution to the field of generative modeling, with the prospect of addressing the longstanding challenge of training stability in VAEs, thereby opening avenues for more efficient and effective model development and application.