Mixed Precision Neural Architecture Search for Energy Efficient Deep Learning

Gong, Chengyue; Jiang, Zixuan; Wang, Dilin; Lin, Yibo; Liu, Qiang; Pan, David Z.

doi:10.1109/iccad45719.2019.8942147

Cited by 53 publications

(43 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Energy is often profiled by NVIDIA's given hardware platform profilers, such as nvprof. We can formalize energy either as peak or average power usage, and the two metrics are used by various HW-NAS works, including [Hsu et al, 2018;Gong et al, 2019]. Area.…”

Section: Hardware Metrics Collectionmentioning

confidence: 99%

Hardware-Aware Neural Architecture Search: Survey and Taxonomy

Benmeziane

Maghraoui²,

Ouarnoughi

et al. 2021

Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence

View full text Add to dashboard Cite

There is no doubt that making AI mainstream by bringing powerful, yet power hungry deep neural networks (DNNs) to resource-constrained devices would required an efficient co-design of algorithms, hardware and software. The increased popularity of DNN applications deployed on a wide variety of platforms, from tiny microcontrollers to data-centers, have resulted in multiple questions and challenges related to constraints introduced by the hardware. In this survey on hardware-aware neural architecture search (HW-NAS), we present some of the existing answers proposed in the literature for the following questions: "Is it possible to build an efficient DL model that meets the latency and energy constraints of tiny edge devices?", "How can we reduce the trade-off between the accuracy of a DL model and its ability to be deployed in a variety of platforms?". The survey provides a new taxonomy of HW-NAS and assesses the hardware cost estimation strategies. We also highlight the challenges and limitations of existing approaches and potential future directions. We hope that this survey will help to fuel the research towards efficient deep learning.

show abstract

Section: Hardware Metrics Collectionmentioning

confidence: 99%

Hardware-Aware Neural Architecture Search: Survey and Taxonomy

Benmeziane

Maghraoui²,

Ouarnoughi

et al. 2021

Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence

View full text Add to dashboard Cite

show abstract

“…While the previously discussed work [10,11,25,28,39] takes a more systematic approach, others [15,35,36,38] leverage machine learning to address the challenge of mixed precision's large search space. [15,36] are more heavy-handed in their approaches. Hardware-Aware Quantization (HAQ) [36] uses reinforcement learning that Uniform precision.…”

Section: Mixed Precision Quantizationmentioning

confidence: 99%

“…takes hardware simulator results on latency and energy into account to satisfy the given resource constraints to find the optimal mixed precision bit width. [15] searches for the optimal combination of NN architecture and quantization scheme by adding quantization as a search parameter during Neural Architecture Search (NAS). NAS involves automating the creation and search for new NN topological structures that outperform hand-designed ones.…”

Section: Mixed Precisionmentioning

confidence: 99%

Neural Network Quantization for Efficient Inference: A Survey

Weng¹

2021

Preprint

View full text Add to dashboard Cite

As neural networks have become more powerful, there has been a rising desire to deploy them in the real world; however, the power and accuracy of neural networks is largely due to their depth and complexity, making them difficult to deploy, especially in resource-constrained devices. Neural network quantization has recently arisen to meet this demand of reducing the size and complexity of neural networks by reducing the precision of a network. With smaller and simpler networks, it becomes possible to run neural networks within the constraints of their target hardware. This paper surveys the many neural network quantization techniques that have been developed in the last decade. Based on this survey and comparison of neural network quantization techniques, we propose future directions of research in the area.

show abstract

“…It has been estimated that energy use by information and communication technology could reach 20% of all electricity we use by 2025 and account for one twentieth of all carbon emissions (Gelenbe & Caseau, 2015; Le Page, 2018). The next‐generation exaflop computing systems could consume as much as 100 MW of energy, equivalent to the output of a small power station (Gong et al, 2019; Marks, 2016).…”

Section: Introductionmentioning

confidence: 99%

“…The designers of computing systems have now recognized that the efficiency of energy use is of primary concern. Computers are now being designed with some brain‐like characteristics such as mixed precision processing (Gong et al, 2019; Marks, 2016), parallel processing (Eddy & Allman, 2000; Gent, 2018), self‐organizing computation (Larkin, Li, Liakh, & Messer, 2020), reduced code precision (Gent, 2018) and sparse coding systems (Olshausen & Field,).…”

Section: Introductionmentioning

confidence: 99%

Computational intelligence within a resource budget: The case of the honey bee (Apis mellifera) swarm

Foss

2020

Syst Res Behav Sci

View full text Add to dashboard Cite

This study investigated the resource budget available to a typical honey bee swarm as it gathered information and decided which was the best of some 13–34 potential nest sites to move into and become its new home. After a brief description of the decision‐making mechanisms involved, the utilization of the swarm's finite energy, memory and carrier resources is described. It was found that each individual scout decision maker only carried enough energy reserves to provide for 2.6 days of visits to evaluate distant sites and, effectively, only enough memory space to remember the location and quality of approximately one site. Scouts, who only formed 4.4% of the swarm bees, consumed 22.8% of the swarm's energy budget, showing that their best‐of‐N computations were energetically very expensive. A typical swarm of 12,000 bees only contained enough energy reserves to last for 13.5 days after which foraging had to restart.

show abstract

Mixed Precision Neural Architecture Search for Energy Efficient Deep Learning

Cited by 53 publications

References 23 publications

Hardware-Aware Neural Architecture Search: Survey and Taxonomy

Hardware-Aware Neural Architecture Search: Survey and Taxonomy

Neural Network Quantization for Efficient Inference: A Survey

Computational intelligence within a resource budget: The case of the honey bee (Apis mellifera) swarm

Contact Info

Product

Resources

About