Leveraging AutoML for Sustainable Deep Learning: A Multi-Objective HPO Approach on Deep Shift Neural Networks
Keywords: Deep Learning, AutoML, Green AutoML, Sustainability, Multi-Objective Optimization, Multi-Fidelity Optimization, Deep Shift Neural Networks
TL;DR: We use multi-fidelity multi-objective optimization to investigate the structure of Deep Shift Neural Networks by analysing the tradeoffs in terms of emissions and accuracy.
Abstract: Deep Learning (DL) has advanced various fields by extracting complex patterns from large datasets.
However, the computational demands of DL models pose environmental and resource challenges. Deep Shift Neural Networks (DSNNs) improve the situation by leveraging shift operations to reduce computational complexity at inference.
Compared to common DNNs, DSNNs are still less well understood and less well optimized.
By leveraging AutoML techniques, we provide valuable insights into the potential of DSNNs and how to design them in a better way.
Following the insights from common DNNs, we propose to leverage the full potential of DSNNs by means of AutoML techniques.
We study the impact of hyperparameter optimization (HPO) on maximizing DSNN performance while minimizing resource consumption.
Since we consider complementary objectives such as accuracy and energy consumption, we combine state-of-the-art multi-fidelity (MF) HPO with multi-objective optimization to find a set of Pareto-optimal trade-offs on how to design DSNNs.
Our approach led to significantly better configurations of DSNNs regarding loss and emissions compared to default DSNNs. This includes simultaneously increasing performance by about 20% and reducing emissions by about 10%.
Investigating the behavior of quantized networks in terms of both emissions and accuracy, our experiments reveal surprising model-specific trade-offs, yielding the greatest energy savings.
For example, in contrast to common expectations, selectively quantizing smaller portions of the network with low precision is optimal while retaining or improving performance.
We corroborated these findings across multiple backbone architectures, highlighting important nuances in quantization strategies and offering an automated approach to balancing energy efficiency and model performance.
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10325
Loading