- author = "Alexander E. I. Brownlee and Jason Adair and Saemundur O. Haraldsson and John Jabbo",
- title = "Exploring the Accuracy -- Energy Trade-off in Machine Learning",
- booktitle = "GI @ ICSE 2021",
- year = "2021",
- month = "30 " # may,
- editor = "Justyna Petke and Bobby R. Bruce and Yu Huang and Aymeric Blot and Westley Weimer and W. B. Langdon",
- publisher = "IEEE",
- address = "internet",
- pages = "11--18",
- keywords = "genetic algorithms, genetic programming, genetic improvement, AI, ML, PyRAPL, NSGA-II jMetalPy",
- isbn13 = "978-1-6654-4466-8/21",
-
URL = "
https://geneticimprovementofsoftware.com/paper_pdfs/gi2021icse/brownlee_gi-icse_2021.pdf",
-
video_url = "
https://www.youtube.com/watch?v=bap72BF9vZw&list=PLI8fiFpB7BoKDaxvS7SQp0iA7fN7rrvDD&index=7",
-
video_url = "https://www.youtube.com/watch?v=tgRV-AsVYko&list=PLI8fiFpB7BoKDaxvS7SQp0iA7fN7rrvDD&index=8",
-
video_url = "https://www.youtube.com/watch?v=tB3IWHT4FT4&list=PLXTjhGKkSnI-se7uQneCX-pEDiDrQ7TIS&index=4",
-
DOI = "
doi:10.1109/GI52543.2021.00011",
- size = "8 pages",
- abstract = "Machine learning accounts for considerable global electricity demand and resulting environmental impact, as training a large deep-learning model produces 284000 kgs of the greenhouse gas carbon dioxide. In recent years, search-based approaches have begun to explore improving software to consume less energy. Machine learning is a particularly strong candidate for this because it is possible to trade off functionality (accuracy) against energy consumption, whereas with many programs functionality is simply a pass-or-fail constraint. We use a grid search to explore hyperparameter configurations for a multilayer perceptron on five classification data sets, considering trade-offs of classification accuracy against training or inference energy. On one data set, we show that 77 percent of energy consumption for inference can saved by reducing accuracy from 94.3 percent to 93.2 percent. Energy for training can also be reduced by 30-50 percent with minimal loss of accuracy. We also find that structural parameters like hidden layer size is a major driver of the energy-accuracy trade-off, but there is some evidence that non-structural hyperparameters influence the trade-off too. We also show that a search-based approach has the potential to identify these trade-offs more efficiently than the grid search.",
-
notes = "data etc at http://hdl.handle.net/11667/173
Energy cost versus accuracy tradeoff, eg for tanh, relu, logistic regression. Hidden layer size (only on glass dataset). Solver (high impact). Model size, esp for online inference (eg for IoT).
Video tB3IWHT4FT4 Alexander E.I. Brownlee.
15:14 Discussion chair: Yu Huang Q: Alexandre Bergel, A: Alexander E.I. Brownlee, pyRAPL easy.
16:01 Q: Westley Weimer, A: learning algorithm
14:34 W. B. Langdon A: pyRAPL v. nanoseconds v. Markus Wagner energy measurements in Joules. 19:00 Approximate computing
19:37 Yu Huang, A: python (SciKit Learn) v. Weka.
20:35 Yu Huang problems, A: noise. SE for AI.
part of \cite{Petke:2021:ICSEworkshop} http://geneticimprovementofsoftware.com/events/icse2021.html",
