Search papers, labs, and topics across Lattice.
This study utilizes AutoML techniques to optimize Deep Shift Neural Networks (DSNNs) for image classification by balancing accuracy and energy consumption through a multi-objective hyperparameter optimization (HPO) framework. The research reveals that DSNNs can achieve a 20% increase in performance while simultaneously reducing emissions by over 60% compared to default configurations. Notably, the findings challenge conventional wisdom about quantization strategies, demonstrating that selectively quantizing smaller portions of the network can yield optimal energy savings without sacrificing performance.
Achieving a 20% boost in performance while slashing emissions by over 60% reveals the untapped potential of Deep Shift Neural Networks in sustainable AI.
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.
