LG AI Research 126

ICLR 2021 is a world-class AI conference covering the representation, learning, and optimization of deep learning models, and this event introduced a virtual conference format using Metaverse technology, providing a unique experience in poster sessions and elsewhere.

Recent research trends are being led by Generative Learning, Meta Learning, and Self-supervised Learning. In particular, beyond existing VAE, GAN, and Normalizing Flow, research on generative models that can solve various problems through data distribution estimation and domain alignment is being actively pursued.

Among these, Score-based Generative Modeling is drawing attention for providing more stable training and superior performance than GAN. In particular, Yang Song's paper treats the process of adding noise to data through a Stochastic Differential Equation (SDE) as a continuous-time stochastic process, and was selected as a Best Paper at ICLR 2021.

This paper unifies the existing SMLD (Score Matching with Langevin Dynamics) and DDPM (Denoising Diffusion Probabilistic Modeling) as VE SDE and VP SDE within the SDE framework. It also proposes a new sampling technique called PC Sampler, which has a Predictor-Corrector structure, enabling more sophisticated image generation.