AI BriefingLG AI Research 119
Key point
Introduces the latest trends in generative models, including the Score-based models and VAEBM that drew attention at ICLR 2021.
Details
ICLR 2021 was held as a virtual conference using a metaverse format, with Generative model, Meta learning, and Self-supervised learning emerging as major keywords.
In particular, in the field of generative models, beyond existing VAE, GAN, and Normalizing Flow, Score-based model and Energy-based model(EBM) are establishing themselves as a new trend by showing meaningful performance.
The featured paper Score-based generative modeling views data as a stochastic process over continuous time $t$, and unified the model through SDE(Stochastic Differential Equation). The authors proposed the PC sampler algorithm, achieving high-resolution image generation and SOTA performance.
Another study, VAEBM, is a model that combines VAE and EBM. To solve the problem of high computational cost in MCMC-based sampling, it combined a VAE component, and through a Two-stage training algorithm it improved optimization efficiency, securing fast sampling speed.