Neighborhood-Based Stopping Criterion for Contrastive Divergence

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Neuroscience

Background:

• Restricted Boltzmann Machines (RBMs) are key for generative modeling in unsupervised learning.
• Contrastive Divergence (CD) is a common RBM training algorithm, approximating log-likelihood gradients.
• Reconstruction error is a frequently used, yet potentially unreliable, stopping criterion for CD training.

Purpose of the Study:

• To address the limitations of reconstruction error as a stopping criterion for RBM training.
• To propose a novel, computationally inexpensive stopping criterion for CD learning.
• To improve the reliability of RBM training by identifying optimal stopping points.

Main Methods:

• Investigated the limitations of reconstruction error for determining RBM training convergence.
• Proposed a new stopping criterion for CD learning based on information from neighboring states.
• Evaluated the effectiveness of the proposed method as an alternative to computationally expensive techniques like annealed importance sampling.

Main Results:

• Demonstrated that reconstruction error can be a misleading indicator of model convergence.
• The proposed method offers a simple and computationally efficient alternative for stopping CD learning.
• The new criterion leverages information from neighboring states to better estimate optimal training points.

Conclusions:

• Reconstruction error is an inadequate stopping criterion for RBM training due to its non-monotonic relationship with log-likelihood.
• A novel stopping criterion utilizing neighboring state information provides a practical and efficient alternative.
• This approach enhances the reliability and efficiency of unsupervised learning with RBMs.