Sampling Method of MCMC Algorithm after Neural Network Optimization in Visual Recognition Field
XiaoJun Zhan, Ren Zhang, YuXia Wang, Jun Tian · 2024
Complex probabilistic models often require computationally intractable high-dimensional integrals in machine learning and Bayesian inference. There are many models that provide approximations to Bayesian probabilistic models, one of which is the Markov Chain Monte Carlo (MCMC). Based on the traditional MCMC model, a Neural Networks Langevin Monte Carlo (NNLMC) model for Langevin dynamics and neural network optimization is proposed. The model reconstructs and optimizes the calculation method and loss function in the traditional algorithm, and improves the convergence speed of the model. In order to verify the convergence speed and efficiency of the sampler, this paper compares the proposed model with the existing Hamiltonian Monte Carlo (HMC), Metropolis-Adjusted Langevin Algorithm (MALA), and Magnetic Hamiltonian Monte Carlo (MHMC) model in terms of autocorrelation, maximum mean difference, the effective length of the sample and the consumption time. Experimental results show that the NNLMC model can efficiently sample from the target distribution.