Revisiting LARS for Large Batch Training Generalization of Neural Networks

LARS and LAMB have emerged as prominent techniques in Large Batch Learning (LBL) to ensure training stability in AI. Convergence stability is a challenge in LBL, where the AI agent usually gets trapped in the sharp minimizer. To address this challenge, warm-up is an efficient technique, but it lacks a strong theoretical foundation. Specifically, the warm-up process often reduces gradients in the early phase, inadvertently preventing the agent from escaping the sharp minimizer early on. In light of this situation, we conduct empirical experiments to analyze the behaviors of LARS and LAMB with and without a warm-up strategy. Our analyses give a comprehensive insight into the behaviors of LARS, LAMB, and the necessity of a warm-up technique in LBL, including an explanation of their failure in many cases. Building upon these insights, we propose a novel algorithm called Time Varying LARS (TVLARS), which facilitates robust training in the initial phase without the need for warm-up. A configurable sigmoid-like function is employed in TVLARS to replace the warm-up process to enhance training stability. Moreover, TVLARS stimulates gradient exploration in the early phase, thus allowing it to surpass the sharp minimizes early on and gradually transition to LARS and achieving robustness of LARS in the latter phases. Extensive experimental evaluations reveal that TVLARS consistently outperforms LARS and LAMB in most cases, with improvements of up to 2% in classification scenarios. Notably, in every case of self-supervised learning, TVLARS dominates LARS and LAMB with performance improvements of up to 10%.