Leveraging AutoML for Sustainable Deep Learning: A Multi-Objective HPO Approach on Deep Shift Neural Networks
2026-06-22 • Machine Learning
Machine Learning
AI summaryⓘ
The authors studied Deep Shift Neural Networks (DSNNs), which are specialized deep learning models that use shift operations to save computing power, especially during image classification tasks. They used automated machine learning (AutoML) to find better DSNN designs that balance accuracy with lower energy use, achieving up to 20% better performance and 60% less emissions compared to standard setups. Surprisingly, their experiments showed that using lower precision (quantization) on only parts of the network can save a lot of energy without losing accuracy. They tested their methods on different network types and offered a way to automatically optimize DSNNs for both power efficiency and accuracy.
Deep LearningDeep Shift Neural NetworksAutoMLImage ClassificationHyperparameter OptimizationMulti-objective OptimizationQuantizationEnergy EfficiencyComputational ComplexityPareto Optimal
Authors
Leona Hennig, Marius Lindauer
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) present a solution 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. We focus on image classification, a core task in computer vision, especially in low-resource environments. Since we consider complementary objectives such as accuracy and energy consumption, we combine state-of-the-art multi-fidelity (MF) hyperparameter optimization (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, in some cases by more than 60%. 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, quantizing smaller portions of the network with low precision can be optimal with respect to energy consumption 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.