An Empirical Comparison of Joint-Training and Pre-Training for Domain-Agnostic Semi-Supervised Learning Via Energy-Based Models

Some semi-supervised learning (SSL) methods heavily rely on domain-specific data augmentations. Recently, semi-supervised learning (SSL) via energy-based models (EBMs) has been studied and is attractive from the perspective of being domain-agnostic, since it inherently does not require data augmentations. There exist two different methods for EBM based SSL - joint-training and pre-training. Joint-training estimates the joint distribution of observations and labels, while pre-training is taken over observations only and followed by fine-tuning. Both joint-training and pre-training are previously known in the literature, but it is unclear which one is better when evaluated in a common experimental setup. To the best of our knowledge, this paper is the first to systematically compare joint-training and pre-training for EBM-based for SSL, by conducting a suite of experiments across a variety of domains such as image classification and natural language labeling. It is found that joint-training EBMs outperform pre-training EBMs marginally but nearly consistently, presumably because the optimization of joint-training is directly related to the targeted task, while pre-training does not.