Follow Hamiltonian Leader: An Efficient Energy-Guided Sampling Method

Yunfei Teng, Yao Li, Kai Chen, Sixin Zhang, Di He, Qiwei Ye

Keywords: Sampling, Hamiltonian Monte Carlo, Monte Carlo Markov Chain

Abstract: Our research underscores the value of leveraging zeroth-order information for addressing sampling challenges, particularly when first-order data is unreliable or unavailable. In light of this, we have developed a novel parallel sampling method that incorporates a leader-guiding mechanism. This mechanism forges connections between multiple sampling instances via a selected leader, enhancing both the efficiency and effectiveness of the entire sampling process. Our experimental results demonstrate that our method markedly expedites the exploration of the target distribution and produces superior quality outcomes compared to traditional sampling techniques. Furthermore, our method also shows greater resilience against the detrimental impacts of corrupted gradients as intended.

Primary Area: Probabilistic methods (for example: variational inference, Gaussian processes)

Submission Number: 14999