Stable Parallel Algorithms for Interdisciplinary Computer-Based Online Education with Real Problem Scenarios for STEM Education

Abstract

In this paper, we analyse and study the interdisciplinary style of stable parallel algorithms for online computer education and real problem scenarios for STEM education. Under the guidance of the STEM concept, the project theme is designed based on the principles of project design, and the interdisciplinary knowledge points related to the theme are determined. Based on the project theme, the teaching design model of the STEM-based robotics project was constructed. The analysis of the results showed that the STEM-based middle-school robotics project not only increased students' interest in robotics but also promoted their creativity, problem-solving, teamwork, and communication skills, as well as their learning of other subjects. In this paper, an adaptive overflow-aware loss amplification strategy is proposed, which effectively alleviates the training nonconvergence problem caused by gradient overflow in mixed accuracy training. Also, this paper demonstrates that data-parallel training under multiple machines and multiple cards should use local batch normalization, which is particularly effective in accelerating neural networks containing more batch normalization layers.