Fuzzy-Based Frugality Score

The machine learning community is increasingly aware of the environmental impact of data processing, especially since the emergence of deep learning models. The increase power consumption to train and deploy deep leanring models has set the machine learning community to consider efficiency as a key aspect of their designs. However, recent studies [1, 2, 3] and regulations [4] tend to shift the focus towards a different concept: frugality.

This dashboard implements a Mamdani fuzzy inference system (FIS) [5] to compute a frugality score based on performance and energy consumption. Fuzzy logic is particularly suitable for multi-criteria decision-making tasks involving uncertainty and human subjectivity [6, 7].
As the performance and energy consumption of a system can be considered as human subjective concepts, a fuzzy-based approach permits to formalize the reasoning process of the user and to compute an interpretable frugality score.

Frugality

Context-aware sustainability

Frugality is context specific and aims to optimize the performance and energy consumption based on real-life interpretable rules. As such, frugality questions the objectives of an application, and can lead to more sustainable, realistic, and user-centric designs.

Efficiency

Performance-to-energy ratio

Efficiency focuses on the trade-off between performance and resource use, here energy consumption. Widely present in the literature, this concept takes little to no consideration of the context of use.

How to use this dashboard?

1
System type"NON-ML" for one energy consumption along performance, or "ML" for energy consumptions in both training and inference.
2
Performance metricSelect the performance metric for classical objective parametrisation from the literature.
3
HardwareSelect CPU and GPU model, with power information fetched from the literature (customisable).
4
ThresholdsSelect time thresholds for the definition of the membership functions associated with the energy consumption. (See page ①)
5
Your valuesEnter your measured performance and energy consumption(s)
6
Rule definitionSelect your wished association rules on page ②.
7
ResultsSelect defuzzification mode and view the score on page ③ along with rule activation outputs.

You can toggle DEMO MODE in the sidebar to choose an execution scenario applied on the data analysed within the linked paper.
1. Training of non-neural-network based classification methods on MNIST
2. Training of benchmark neural networks on ImageNet from scratch
3. Training of benchmark neural networks on CIFAR100 from scratch
4. Training of benchmark neural networks on CIFAR100 pretrained on ImageNet
5. Training of rectangular-shaped multi-layer perceptrons with different depths on MNIST (from BUTTER-E [8])

References

[1] N. Girdhar, A. Raj, D. Sharma, V. Singh, A. Doucet, and M. Renz, “A comprehensive review of frugal artificial intelligence: challenges, applications, and the road to sustainable AI,” Soft Computing, vol. 29, pp. 4823–4856, July 2025.
[2] R. Basu, P. Banerjee, and E. Sweeny, “Frugal Innovation: Core Competencies to Address Global Sustainability,” Journal of Management for Global Sustainability, vol. 1, Dec. 2013.
[3] IPCC, “Summary for policymakers,” in Climate Change 2022: Mitigation of Climate Change. Contribution of Working Group III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (P. Shukla, J. Skea, R. Slade, A. A. Khourdajie, R. van Diemen, D. McCollum, M. Pathak, S. Some, P. Vyas, R. Fradera, M. Belkacemi, A. Hasija, G. Lisboa, S. Luz, and J. Malley, eds.), Cambridge, UK and New York, NY, USA: Cambridge University Press, 2022.
[4] AFNOR, “AFNOR SPEC 2314: Référentiel général pour l'IA frugale – Mesurer et réduire l'impact environnemental de l'IA,” AFNOR Éditions, 2024. [5] E. H. Mamdani and S. Assilian, “An experiment in linguistic synthesis with a fuzzy logic controller,” International Journal of Man-Machine Studies, vol. 7, pp. 1–13, Jan. 1975.
[6] F. A. Lootsma, Fuzzy Logic for Planning and Decision Making. Springer Science & Business Media, Mar. 2013. Google-Books-ID: 7GXmB-wAAQBAJ.
[7] L. Chen, G. Duan, S. Wang, and J. Ma, “A Choquet integral based fuzzy logic approach to solve uncertain multi-criteria decision making problem,” Expert Systems with Applications, vol. 149, p. 113303, July 2020.s
[8] C. Tripp, J. Perr-Sauer, E. Bensen, J. Gafur, A. Nag, and A. Purkayastha, “Butter-e - energy consumption data for the butter empirical deep learning dataset.” Open Energy Data Initiative (OEDI), National RenewableEnergy Laboratory, https://doi.org/10.25984/2329316, 2022. Accessed: 2026-01-13.

Performance membership functions
Energy membership functions (train)
Score output membership functions
Rule base matrix rows = energy · cols = performance · cells = score output
E ↓   P → P = lowP = mediumP = high

Frugality score
/ 100
Energy memberships
Low
Medium
High
Performance memberships
Low
Medium
High
Output fuzzy set activations
Aggregated output & defuzzification
Rule activations