On the Understandability of Machine Learning Practices in Deep Learning and Reinforcement Learning Based Systems

作     者：Ntentos, Evangelos Warnett, Stephen John Zdun, Uwe 

作者机构：Faculty of Computer Science Research Group Software Architecture University of Vienna Vienna Austria UniVie Doctoral School Computer Science DoCS Faculty of Computer Science University of Vienna Vienna Austria 

出 版 物：《SSRN》 

年 卷 期：2024年

核心收录：

主　　题：Predictive analytics 

摘      要：Machine learning (ML) utilizes diverse algorithms for predictive analytics. Deep Learning (DL) employs neural networks for intricate problem-solving, while Reinforcement Learning (RL) tackles sequential decision-making challenges. Best practices like transfer learning and checkpoints address issues with vast datasets and generalization to new *** ML systems solely from source code is difficult, particularly for novice developers. Our study hypothesizes that incorporating ML system diagrams detailing workflows and practices can improve comprehension in system design tasks. We anticipate this enhancement will positively impact correctness and duration of tasks, revealing varying participant *** findings suggest integrating semi-formal ML system diagrams alongside source code enhances task correctness in DL sub-tasks. However, differences in performance between participants who received only source code and those with diagrams were less pronounced in RL sub-tasks. This implies varied efficacy across learning settings, highlighting the need for deliberate diagram usage based on practitioners goals. © 2024, The Authors. All rights reserved.