Search

Automated hyperparameter optimization (HPO) is a crucial and time consuming part in the automatic generation of efficient machine learning models. Previous studies could be classified into two major categories in terms of reducing training overhead: (1) sampling a promising hyperparameter configuration and (2) pruning non-promising configurations. These adaptive sampling and resource scheduling are combined to reduce cost, increasing the number of evaluations done on more promising configurations to find the best model in a given time. That is, these strategies are preferred to identify the best-performing models at an early stage within a certain deadline. Although these time and resource constraints are significant for designing HPO strategies, previous studies only focused on parallel exploration efficiency using resource awareness. In this study, we propose a novel diversification strategy for HPO, which exploits the dynamic hyperparameter space allocation for a sampler according to the remaining time budget. We provide a simple yet effective method to accelerate the maturity of the sampler that is independent of the sampling algorithm. Compared to previous resource awareness solutions, our solution achieves better performance via both time and resource awareness. We demonstrate the performance gains of our solution on several well-known HPO benchmarks. Furthermore, we implement them to our high performance computing AI convergence platform. Considering the different types of users, both a fully automated HPO service based on graphic processing unit (GUI) interfaces and HPO job management via python application programming interface (API) on the Jupyter lab are served on the platform, publicly. [ABSTRACT FROM AUTHOR]

With the increase in research on AI (Artificial Intelligence), the importance of DL (Deep Learning) in various fields, such as materials, biotechnology, genomes, and new drugs, is increasing significantly, thereby increasing the number of deep-learning framework users. However, to design a deep neural network, a considerable understanding of the framework is required. To solve this problem, a GUI (Graphical User Interface)-based DNN (Deep Neural Network) design tool is being actively researched and developed. The GUI-based DNN design tool can design DNNs quickly and easily. However, the existing GUI-based DNN design tool has certain limitations such as poor usability, framework dependency, and difficulty encountered in changing GUI components. In this study, a deep learning algorithm that solves the problem of poor usability was developed using a template to increase the accessibility for users. Moreover, the proposed tool was developed to save and share only the necessary parts for quick operation. To solve the framework dependency, we applied ONNX (Open Neural Network Exchange), which is an exchange standard for neural networks, and configured it such that DNNs designed with the existing deep-learning framework can be imported. Finally, to address the difficulty encountered in changing GUI components, we defined and developed the JSON format to quickly respond to version updates. The developed DL neural network designer was validated by running it with KISTI's supercomputer-based AI Studio. [ABSTRACT FROM AUTHOR]

Summary: With the recent emergence of new paradigm, ie, open science and big data, the need for data sharing and collaboration is becoming important in the computational science field as well. The EDISON‐DATA platform aims to provide services that computational simulation data can easily published, preserved, shared, reused, discovered, and analyzed. First, this paper analyzed computational science platform‐related issues, obtained during the development of the EDISON‐DATA platform, regarding the sharing and reusing of the computational science data. These issues include data complexity, diversity, reliability, heterogeneity, etc. To solve the above issues and support data analysis in an efficient and integrated manner, this study proposes various ideas used in the EDISON‐DATA platform. First, we suggested an automated preprocessing framework to handle the complexity of computational science data. Second, to solve the diversity issue, we presented ways to develop preprocessing logic and data presentation logic customized for each data type. Third, to improve the reliability of computational science data, some quality control and provenance management techniques were presented. Fourth, we proposed a way to manage related data in groups. Fifth, to solve data heterogeneity problem and to analyze data in an integrated way, we let the preprocessing framework to use controlled vocabularies to express descriptive metadata. Lastly, we demonstrated feasibility and usability of the proposed ideas in this paper by presenting a case study of building a research portal service in the materials field based on the EDISON‐DATA platform. [ABSTRACT FROM AUTHOR]