情報処理学会 第86回全国大会 会期:2024年3月15日~17日

7C-02
Adaptive HMC: Improve Generation Quality of Score-based Generative Model
○韓  宇,中村和幸(明大)
Score-based Generative Modelss (SGMs) is increasingly gaining attention for its ease of training. Langevin MCMC is commonly used in SGMs, but it introduces inefficient random walks. Hamiltonian Monte Carlo (HMC) is a deterministic MCMC method that utilizes an auxiliary variable scheme. Suppressing the random walk is a significant feature of HMC. Therefore, in this study, we employ HMC during the generation process of SGMs. We will present the experimental results of the proposed approach.