MMD Finetuning Calibrates Generative Models to Feature Distributions

This paper presents kernel Calibrating Generative Models (kCGM), a novel approach to address the issue of distributional miscalibration in generative models. While generative models can produce plausible individual samples, their overall feature distribution might significantly deviate from a desired target set. Direct finetuning on the target set often leads to overfitting and compromises the validity of generated samples. kCGM tackles this by minimizing the Maximum Mean Discrepancy (MMD) between the feature distributions of generated and target data. It employs an unbiased score-function estimator for MMD minimization, coupled with KL regularization to ensure the finetuned model remains close to its pretrained state. The effectiveness of kCGM is demonstrated across various tasks, including drug-like molecule generation, protein generation, and DNA generation. For instance, in generating antibiotics, kCGM significantly improves target feature matching while maintaining chemical validity, a challenge for direct finetuning. The method is adaptable to different generative model architectures, including autoregressive, continuous-space diffusion, and discrete diffusion models, using only feature-level supervision.