Task-Specific Automation in Deep Learning Processes

Abstract:

Recent advances in deep learning facilitate the training, testing, and deployment of models through so-called pipelines. Those pipelines are typically orchestrated with general-purpose machine learning frameworks (e.g., Tensorflow Extended), where developers manually call the single steps for each task-specific application. The diversity of task- and technology-specific requirements in deep learning projects increases the orchestration effort. There are recent advances to automate the orchestration with machine learning, which are however, still immature and do not support task-specific applications. Hence, we claim that partial automation of pipeline orchestration with respect to specific tasks and technologies decreases the overall development effort. We verify this claim with the ALOHA tool flow, where task-specific glue code is automated. The gains of the ALOHA tool flow pipeline are evaluated with respect to human effort, computing performance, and security.

Access paper here

BibTeX:

@inproceedings{10.1007/978-3-030-87101-7_16,
author = {Buchgeher, Georg and Czech, Gerald and Ribeiro, Adriano Souza and Kloihofer, Werner and Meloni, Paolo and Busia, Paola and Deriu, Gianfranco and Pintor, Maura and Biggio, Battista and Chesta, Cristina and Rinelli, Luca and Solans, David and Portela, Manuel},
editor = {Kotsis, Gabriele and Tjoa, A. Min and Khalil, Ismail and Moser, Bernhard and Mashkoor, Atif and Sametinger, Johannes and Fensel, Anna and Martinez-Gil, Jorge and Fischer, Lukas and Czech, Gerald and Sobieczky, Florian and Khan, Sohail},
title = {Task-Specific Automation in Deep Learning Processes},
booktitle = {Database and Expert Systems Applications - DEXA 2021 Workshops},
isbn = {978-3-030-87101-7},
pages = {159–169},
address = {Cham},
publisher = {Springer International Publishing},
year = {2021},
url = {https://link.springer.com/chapter/10.1007/978-3-030-87101-7_16}
}