Reproduce “Denoising Likelihood Score Matching for Conditional Score-based Data Generation” on OpenResearch

About this paper

Chen-Hao Chao, Wei-Fang Sun, Bo-Wun Cheng, Yi-Chen Lo, Chia-Che Chang, Yu-Lun Liu, Yu-Lin Chang, Chia-Ping Chen and 1 more

Denoising Likelihood Score Matching (DLSM) proposes a new training objective for classifiers in conditional score-based generative models, accurately estimating likelihood scores to resolve the 'score mismatch issue.' This method achieves superior fidelity, diversity, and class consistency in generated samples compared to prior approaches on datasets like CIFAR-10 and CIFAR-100.

Key insights

• Standard cross-entropy loss is inadequate for training classifiers to accurately estimate the gradients of the log-likelihood (likelihood score), leading to a fundamental 'score mismatch' in conditional score-based models.
• Applying denoising score matching principles directly to the likelihood score estimation problem (DLSM) provides a theoretically sound and practical method to correct this mismatch.
• Combining the explicit DLSM loss with traditional cross-entropy loss yields a synergistic effect, simultaneously achieving high likelihood score accuracy and stable classifier training.

Summary by alphaXiv.