Model Compression and Hardware Acceleration for Neural Networks: A Comprehensive Survey
The DOI matches its Crossref record; 317 of the 334 references it cites were found in a registry or an index, and 1 are not checked yet.
Abstract
Domain-specific hardware is becoming a promising topic in the backdrop of improvement slow down for general-purpose processors due to the foreseeable end of Moore's Law. Machine learning, especially deep neural networks (DNNs), has become the most dazzling domain witnessing successful applications in a wide spectrum of artificial intelligence (AI) tasks. The incomparable accuracy of DNNs is achieved by paying the cost of hungry memory consumption and high computational complexity, which greatly impedes their deployment in embedded systems. Therefore, the DNN compression concept was naturally proposed and widely used for memory saving and compute acceleration. In the past few years, a tremendous number of compression techniques have sprung up to pursue a satisfactory tradeoff between processing efficiency and application accuracy. Recently, this wave has spread to the design of neural network accelerators for gaining extremely high performance. However, the amount of related works is incredibly huge and the reported approaches are quite divergent. This research chaos motivates us to provide a comprehensive survey on the recent advances toward the goal of efficient compression and execution of DNNs without significantly compromising accuracy, involving both the high-level algorithms and their applications in hardware design. In this article, we review the mainstream compression approaches such as compact model, tensor decomposition, data quantization, and network sparsification…
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“but there is no review discussing all three categories which in future will be integrated seamlessly to design effective and efficient DNNs in the new edge intelligence era.”
From Bringing ai to edge: from deep learning’s perspective (2022) · cites this paper as [28]
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 3×
“Specifically, an example of the identity connection used in ResNet [30] [see Fig.”
BackgroundMethodVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×
“In VGG-Nets [27], the interlayer topology is rather simple, where multiple Conv layers are stacked head-to-tail in sequence.”
BackgroundImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 3×
“1078 [20] A.”
BackgroundInfluentialHochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 2×
“…we discuss recent advances in RNNs that aim to improve the execution efficiency by designing compact RNNs in the level of unit or network. a) Unit level: LSTM is designed to improve the long-term dependence over vanilla RNNs by alleviating the issue of gradient explosion or vanishing [18], [46].”
BackgroundMethodLeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
Densely Connected Convolutional Networks
Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.243 · Cited by 48,566
Going deeper with convolutions
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298594 · Cited by 48,562
(2014) · Cited by 33,419
Rethinking the Inception Architecture for Computer Vision
Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.308 · Cited by 32,138
MobileNetV2: Inverted Residuals and Linear Bottlenecks
Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L.-C. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00474 · Cited by 28,755 · cited 35×
“The feature propagation gradually improves the abstraction level from local to global, which finally enables the network decision such as image recognition. a) Spatial correlation to enlarge RF: Compared to MLP, the key feature of CNNs is the local connections with shared weights, which results in high spatial correlat…”
BackgroundMethodInfluential
Research works printed but not linked: 2
- lin, 2013, Network in network, arXiv 1312 4400
- han, 2015, Learning both weights and connections for efficient neural network, Proc Adv Neural Inf Process Syst, 1135
Cited by
Items 301–400 of 1,002
Distributed Edge System Orchestration for Web-based Mobile Augmented Reality Services
Ren, P., Liu, L., Qiao, X., & Chen, J. (2022) · IEEE Transactions on Services Computing · doi:10.1109/tsc.2022.3190375 · Cited by 21
“To meet the accurate and efficient requirements, researchers have recently been focusing on model compression and acceleration [20], [21], but these solutions are usually time-and resources consuming, making it difficult to provide timely services for AR applications.”
BackgroundTHETA: A High-Efficiency Training Accelerator for DNNs With Triple-Side Sparsity Exploration
Lu, J., Huang, J., & Wang, Z. (2022) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2022.3175582 · Cited by 15
Edge-Assisted Real-Time Instance Segmentation for Resource-Limited IoT Devices
Xie, Y., Guo, Y., Mi, Z., Yang, Y., & Obaidat, M. S. (2022) · IEEE Internet of Things Journal · doi:10.1109/jiot.2022.3199921 · Cited by 18
“The studies on this topic have mainly focused on model quantization, layer or tensor fusion, and dynamic memory optimization for spe-cific hardware or kernel specialization for given tasks [31].”
BackgroundSelf-citationTowards designing a hardware accelerator for 3D convolutional neural networks
Khan, F. H., Pasha, M. A., & Masud, S. (2022) · Computers & Electrical Engineering · doi:10.1016/j.compeleceng.2022.108489 · Cited by 15
Xie, H., Liu, Y., Lei, H., Song, T., Yue, G., Du, Y., Wang, T., Zhang, G., & Lei, B. (2022) · Medical Image Analysis · doi:10.1016/j.media.2022.102725 · Cited by 15
Deep-learning-based multi-user framework for end-to-end fiber-MMW communications
Li, Z., Jia, J., Li, G., Dong, B., Shen, W., Huang, C., Shi, J., Chi, N., & Zhang, J. (2023) · Optics Express · doi:10.1364/oe.479500 · Cited by 15
Self-citationIn-Sensor & Neuromorphic Computing Are all You Need for Energy Efficient Computer Vision
Datta, G., Liu, Z., Kaiser, M. A.-A., Kundu, S., Mathai, J., Yin, Z., Jacob, A. P., Jaiswal, A., & Beerel, P. A. (2023) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp49357.2023.10094902 · Cited by 15
“In fact, this energy may actually dominate the total energy incurred by the CV processing, particularly with the range of effective model optimization techniques developed by the CV community [4].”
BackgroundZhong, H., Yu, S. S., Trinh, H. M., Lv, Y., Yuan, R. B., & Wang, Y. (2023) · IEEE Sensors Journal · doi:10.1109/jsen.2023.3332653 · Cited by 16
“To ensure performance and simultaneously achieve lightweight, the concept of model compression has been proposed and widely adopted [17].”
BackgroundNetcast: Low-Power Edge Computing With WDM-Defined Optical Neural Networks
Hamerly, R., Sludds, A., Bandyopadhyay, S., Chen, Z., Zhong, Z., Bernstein, L., & Englund, D. R. (2024) · Journal of Lightwave Technology · doi:10.1109/jlt.2024.3405799 · Cited by 15
“SWaP-constrained hardware [2], [3] and algorithms [4], [5], [6], [7] for DNN edge inference.”
Li, Y., Li, B., Zhang, G., Chen, J., Deng, S., & Zhang, H. (2025) · Remote Sensing · doi:10.3390/rs17203466 · Cited by 14
“However, high-performance IRSTD models often adopt massive connections and complex structures [5], resulting in models with millions of parameters and billions of floating-point operations [6].”
Feature Distillation Siamese Networks for Object Tracking
Huang, H., Liu, G., Zhang, Y., & Xiong, R. (2022) · Applied Soft Computing · doi:10.1016/j.asoc.2022.109912 · Cited by 14
Fast and efficient computing for deep learning-based defect detection models in lightweight devices
Fișne, A., Kalay, A., & Eken, S. (2024) · Journal of Intelligent Manufacturing · doi:10.1007/s10845-024-02487-z · Cited by 15
“Furthermore, the specific computing units to expedite training and inference, results in notable speed enhancements (Deng et al., 2020; Fisne et al., 2023).”
Liu, R., & Zeng, W. (2024) · Structural Health Monitoring · doi:10.1177/14759217241289066 · Cited by 14
Privacy preserving federated learning with convolutional variational bottlenecks
Scheliga, D., Mäder, P., & Seeland, M. (2025) · Cybersecurity · doi:10.1186/s42400-024-00295-9 · Cited by 15
“during collaborative training [41]–[43].”
InfluentialHybrid neural state machine for neural network
Tian, L., Wu, Z., Wu, S., & Shi, L. (2021) · Science China Information Sciences · doi:10.1007/s11432-019-2988-1 · Cited by 14
Self-citationA Marr's Three‐Level Analytical Framework for Neuromorphic Electronic Systems
Guo, Y., Zou, X., Hu, Y., Yang, Y., Wang, X., He, Y., Kong, R., Guo, Y., Li, G., Zhang, W., et al. (2021) · Advanced Intelligent Systems · doi:10.1002/aisy.202100054 · Cited by 16
Self-citationA Deep Learning Model Compression and Ensemble Approach for Weed Detection
Ofori, M. Q., El-Gayar, O., O’Brien, A., & Noteboom, C. (2022) · Hawaii International Conference on System Sciences (HICSS) · doi:10.24251/hicss.2022.138 · Cited by 14
“In summary, as demonstrated by past research [33]–[35], [54], model compression reduces the complexity and resource demands of the DCNN models to allow for faster real-time inference.”
MethodSignal Compression via Neural Implicit Representations
Pistilli, F., Valsesia, D., Fracastoro, G., & Magli, E. (2022) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp43922.2022.9747208 · Cited by 14
Hardware-friendly compression and hardware acceleration for transformer: A survey
Huang, S., Tang, E., Li, S., Ping, X., & Chen, R. (2022) · Electronic Research Archive · doi:10.3934/era.2022192 · Cited by 16
Ghimire, D., & Kim, S. (2022) · Applied Sciences · doi:10.3390/app13010316 · Cited by 15
SignQuery: A Natural User Interface and Search Engine for Sign Languages with Wearable Sensors
Zhou, H., Lu, T., Mckinnie, K., Palagano, J., DeHaan, K., & Gowda, M. (2023) · ACM Annual International Conference on Mobile Computing and Networking (MobiCom) · doi:10.1145/3570361.3613286 · Cited by 16
“To address this challenge, SignQuery intends to utilize model compression techniques [27, 39] to achieve a balance between performance and energy consumption.”
MethodLuo, Y., Luo, Y.-C., & Yu, S. (2021) · IEEE Transactions on Computers · doi:10.1109/tc.2021.3122872 · Cited by 14
Yu, H., & Bro, R. (2021) · Chemometrics and Intelligent Laboratory Systems · doi:10.1016/j.chemolab.2021.104446 · Cited by 15
Raza, S. M., Abidi, S. M. H., Masuduzzaman, M., & Shin, S. Y. (2025) · Neurocomputing · doi:10.1016/j.neucom.2025.131357 · Cited by 13
Liu, F., Zhao, W., Wang, Z., Chen, Y., He, Z., Jing, N., Liang, X., & Jiang, L. (2022) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3489517.3530660 · Cited by 13
Compressing Models with Few Samples: Mimicking then Replacing
Wang, H., Liu, J., Ma, X., Yong, Y., Chai, Z., & Wu, J. (2022) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52688.2022.00078 · Cited by 22
“In order to put deep models into small devices and decrease the latency and memory consumption, network compression [5] is widely used in model deployment.”
Evaluating Spiking Neural Network on Neuromorphic Platform For Human Activity Recognition
Bian, S., & Magno, M. (2023) · International Symposium on Wearable Computers · doi:10.1145/3594738.3611369 · Cited by 13
“This has led to exploring various neural networks compressions strategies, such as pruning and quantization, to have tiny machine learning models that work with a kilobyte of memory and Giga operations per second in the best case [18, 24].”
BackgroundOu, X., Chen, Z., Zhu, C., & Liu, Y. (2023) · Journal of Systems Engineering and Electronics · doi:10.23919/jsee.2023.000159 · Cited by 19
Hassan, H. A. A. (2024) · Journal of Electrical Systems · doi:10.52783/jes.2292 · Cited by 16
Boggust, A., Sivaraman, V., Assogba, Y., Ren, D., Moritz, D., & Hohman, F. (2024) · IEEE Transactions on Visualization and Computer Graphics · doi:10.1109/tvcg.2024.3456371 · Cited by 14
“Model compression encompasses various techniques that reduce the storage space, memory, power, or time required to run an ML model while preserving its original behavior as much as possible [7,8,10,38, 57].”
BackgroundParra, J., Navarro‐Arenas, J., & Sanchis, P. (2025) · Scientific Reports · doi:10.1038/s41598-025-90350-3 · Cited by 14
FORTALESA: Fault-tolerant reconfigurable systolic array for DNN inference
Cherezova, N., Jutman, A., & Jenihhin, M. (2025) · Microprocessors and Microsystems · doi:10.1016/j.micpro.2025.105222 · Cited by 14
“It is commonly used and allows high inputs and outputs reuse [37], [38] and on average has higher performance [33].”
A FeRAM based Volatile/Non-volatile Dual-mode Buffer Memory for Deep Neural Network Training
Luo, Y., Luc, Y.-C., & Yu, S. (2021) · Design, Automation, and Test in Europe · doi:10.23919/date51398.2021.9474180 · Cited by 13
p-Meta: Towards On-device Deep Model Adaptation
Qu, Z., Zhou, Z., Tong, Y., & Thiele, L. (2022) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3534678.3539293 · Cited by 19
“Although weight importance metrics and SGD with sparse gradients have been explored in vanilla training [8, 15, 20, 29], it is unknown (i) how to identify adaptation-critical weights and (ii) whether meta learning is robust to sparse gradients, where the objective is fast adaptation to unseen tasks.”
BackgroundA simple approach for quantizing neural networks
Maly, J., & Saab, R. (2023) · Applied and Computational Harmonic Analysis · doi:10.1016/j.acha.2023.04.004 · Cited by 18
Zou, Y., & Wan, Y. (2024) · IEEE Transactions on Instrumentation and Measurement · doi:10.1109/tim.2024.3381666 · Cited by 14
“These methods are typically categorized into four groups: compact models, tensor decomposition, data quantization, and network sparsification [20].”
MethodCompression strategies and space-conscious representations for deep neural networks
Marinò, G. C., Ghidoli, G., Frasca, M., & Malchiodi, D. (2021) · Proceedings - International Conference on Pattern Recognition/Proceedings/International Conference on Pattern Recognition · doi:10.1109/icpr48806.2021.9412209 · Cited by 13
“The reader interested to a thorough review of these methods can refer for instance to [5], [6].”
BackgroundEciton: Very Low-power Recurrent Neural Network Accelerator for Real-time Inference at the Edge
Chen, J. Y., Jun, S.-W., Hong, S., He, W., & Moon, J. (2023) · ACM Transactions on Reconfigurable Technology and Systems · doi:10.1145/3629979 · Cited by 17
“This result is in part due to the fact recurrent neural networks are significantly more susceptible to accuracy loss from quantization [18].”
BackgroundOptCDU: Optimizing the Computing Data Unit Size for COIN
Wu, H., He, J., Weng, J., Nguyen, G. T. K., Martin Reisslein, B., & Fitzek, F. H. P. (2024) · IEEE Transactions on Network and Service Management · doi:10.1109/tnsm.2024.3452485 · Cited by 12
“…range of available computing options in the network nodes, e.g., computing on general-purpose central processing units (CPUs) as well as computing on specialized hardware [58], [75], [76], [77], [78], [79], whereby tradeoffs between performance and flexibility [80], [81], [82] should be considered.”
Teng, X., Xu, S., Guo, D., Guo, Y., Meng, W., & Zhang, X. (2024) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2024.3454122 · Cited by 13
“Additionally, methods such as quantization and pruning can be employed to enhance inference speed [43].”
MethodA Survey of AI Inference Technologies for On-Device Systems
Wang, W., Li, K., Ji, B., Liu, X., Yu, J., & Wu, Q. (2025) · IEEE Internet of Things Journal · doi:10.1109/jiot.2025.3613691 · Cited by 16
“To solve this, model compression techniques reduce memory usage and computational demands, thereby accelerating inference while maintaining the model’s accuracy [215].”
Aicha, F. B. (2025) · Computing · doi:10.1007/s00607-025-01559-z · Cited by 12
“Optimization Techniques Model quantization, weight pruning, and efficient data representation to minimize memory footprint and improve execution speed [31, 32].”
Khan, M. T., Ding, Y., & Goussetis, G. (2026) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2026.3656642 · Cited by 18
“Deng et al. [42] explore NN compression methods and hardware acceleration platforms, emphasizing the need for algorithm–hardware codesign to support e ffi cient execution of deep NNs. Borel et al. [43] compare linearization methods, including predistortion, in the context of 5G networks.”
Lane Compression: A Lightweight Lossless Compression Method for Machine Learning on Embedded Systems
Ko, Y., Chadwick, A., Bates, D., & Mullins, R. (2021) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3431815 · Cited by 12
“State-of-the-art quantization methods for weights usually quantise W 0 only and also exclude the first and last convolutional layers from quantization to minimise any drop in accuracy [12, 26, 54, 58].”
BackgroundMethodBisimulations for Neural Network Reduction
Prabhakar, P. (2022) · Lecture notes in computer science · doi:10.1007/978-3-030-94583-1_14 · Cited by 12
On Nonlinear Learned String Indexing
Ferragina, P., Frasca, M., Marinò, G. C., & Vinciguerra, G. (2023) · IEEE Access · doi:10.1109/access.2023.3295434 · Cited by 13
A lightweight GRU-based gesture recognition model for skeleton dynamic graphs
Ni, J., Wang, Y., Tang, G., Cao, W., & Yang, S. X. (2024) · Multimedia Tools and Applications · doi:10.1007/s11042-024-18313-w · Cited by 13
“It is also meaningful work to eliminate redundant parts through model compression [32, 33], so that deep learning models can be deployed on devices such as smart phones.”
Talaria: Interactively Optimizing Machine Learning Models for Efficient Inference
Hohman, F., Wang, C., Lee, J., Görtler, J., Moritz, D., Bigham, J. P., Ren, Z., Foret, C., Shan, Q., & Zhang, X. (2024) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3613904.3642628 · Cited by 15
“…support for quantization [66], sparsity [67], and it’s accompanying examples [68]; Google’s quantization extension to Keras called QKeras [31]; Microsoft’s Neural Network Intelligence package and tool [58]; Intel’s Neural Compressor library [47]; and Apple’s MLX framework [38] and DNIKit [90].”
MATADOR: Automated System-on-Chip Tsetlin Machine Design Generation for Edge Applications
Rahman, T., Mao, G., Maheshwari, S., Shafik, R., & Yakovlev, A. (2024) · Design, Automation, and Test in Europe · doi:10.23919/date58400.2024.10546779 · Cited by 15
Mobile Edge Generation-Enabled Digital Twin: Architecture Design and Research Opportunities
Xu, X., Zhong, R., Mu, X., Liu, Y., & Huang, K. (2024) · IEEE Communications Magazine · doi:10.1109/mcom.002.2400224 · Cited by 13
“• Scalable Learning and Compression : There are several candidate techniques for the compression of sketch/seed and GAI sub-models [12].”
MethodZhang, Y., Liang, W., Xu, Z., Jia, X., & Yang, Y. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3540017 · Cited by 12
“( 1) The model compression approach that compresses an inference model (e.g., a DNN) in both the number of neurons and the number of layers to mitigate the computing and memory resource demands of the model, with compromises on the model accuracy and running time [7], [38].”
Progressive principle component analysis for compressing deep convolutional neural networks
Zhou, J., Qi, H., Chen, Y., & Wang, H. (2021) · Neurocomputing · doi:10.1016/j.neucom.2021.01.035 · Cited by 14
Carry-Propagation-Adder-Factored Gemmini Systolic Array for Machine Learning Acceleration
Inayat, K., & Chung, J. (2021) · Electronics · doi:10.3390/electronics10060652 · Cited by 12
“Detailed survey on ML accelerators is available in [18–21].”
MethodLu, A., Peng, X., Luo, Y., Huang, S., & Yu, S. (2021) · ACM Transactions on Design Automation of Electronic Systems · doi:10.1145/3460436 · Cited by 12
Towards Automatic and Agile AI/ML Accelerator Design with End-to-End Synthesis
Zhang, J. J., Agostini, N. B., Song, S., Tan, C., Limaye, A., Amatya, V., Manzano, J., Minutoli, M., Castellana, V. G., Tumeo, A., et al. (2021) · IEEE International Conference on Application-specific Systems, Architectures and Processors (ASAP) · doi:10.1109/asap52443.2021.00040 · Cited by 14
“Furthermore, many hardware templates and improvements designed explicitly for DNN compression and weight sparsity are available [13], [7], [47].”
BackgroundCherupally, S. K., Rakin, A. S., Yin, S., Seok, M., Fan, D., & Seo, J.-S. (2021) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586233 · Cited by 12
“While many works investigated DNN model size reduction [1], DNNs still require a very large number of computation and memory accesses.”
BackgroundZaman, K. S., Reaz, M. B. I., Bakar, A. A. A., Bhuiyan, M. A. S., Arsad, N., Mokhtar, M. H. H., & Ali, S. H. M. (2022) · Engineering Science and Technology an International Journal · doi:10.1016/j.jestch.2022.101153 · Cited by 13
Fresa, A., & Champati, J. P. (2022) · ACM International Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM) · doi:10.1145/3551659.3559044 · Cited by 16
“Thanks to the development of DNN models with reduced computation and storage requirements (at the cost of reduced inference accuracy) and the advancements in the hardware of EDs[11], ML frameworks such as Tensorflow Lite [1] and PyTorch Mobile [2] are now able to support the deployment of DNNs on EDs.”
MethodSoBS-X: Squeeze-Out Bit Sparsity for ReRAM-Crossbar-Based Neural Network Accelerator
Liu, F., Wang, Z., Chen, Y., He, Z., Yang, T., Liang, X., & Jiang, L. (2022) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2022.3172907 · Cited by 12
“In addition, ReRAM crossbar can perform highly parallel vector–matrix multiplication (VMM) by sharing inputs in rows and gathering currents in column using Kirchhoff’s laws [1], [2], [8]–[11].”
BackgroundMethodInfluentialNeural Network Compression for Noisy Storage Devices
Isik, B., Choi, K., Zheng, X., Weissman, T., Ermon, S., Wong, H. ‐. P., & Alaghi, A. (2023) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3588436 · Cited by 14
“…2020; Kundu et al., 2021; Isik et al., 2023; Barnes et al., 2020; Pase et al., 2022), quantization, (Khoram & Li, 2018; Banner et al., 2018; Choi et al., 2020; Fan et al., 2020; Young et al., 2020) and KD (Hinton et al., 2015; Polino et al., 2018; Xie et al., 2020) among others (Deng et al., 2020).”
MethodRethinking Floating Point Overheads for Mixed Precision DNN AcceleratorsPreprint
Abdel-Aziz, H. A., Shafiee, A. R., Shin, J. H., Pedram, A., & Hassoun, J. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2101.11748 · Cited by 17
“DNN accelerators may utilize quantization schemes to reduce DNNs memory footprint and computation time (Deng et al., 2020).”
MethodFull-Analog Reservoir Computing Circuit Based on Memristor With a Hybrid Wide-Deep Architecture
Chen, L., Wang, X., Yang, C., Chen, Z., Zhang, J., & Zeng, Z. (2023) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2023.3334267 · Cited by 14
“Full-Analog Reservoir Computing Circuit Based on Memristor With a Hybrid Wide-Deep Architecture ANNs with complicated structures, the hardware implementation of ANNs increasingly arouses public awareness [6], [7], [8].”
MethodHuang, M., Liu, Y., Zhao, L., & Wang, G. (2023) · Neural Computing and Applications · doi:10.1007/s00521-023-09332-z · Cited by 11
“Nowadays, the model compression methods include two main categories [5]: pruning and quantization.”
Sharma, S. K., Kang, B., Kidambi, N. V., & Mukhopadhyay, S. (2024) · IEEE Sensors Journal · doi:10.1109/jsen.2024.3382479 · Cited by 13
“Prior works have demonstrated the potential of hardware-aware quantization strategies [24].”
Exploring the Boundaries of On-Device Inference: When Tiny Falls Short, Go Hierarchical
Behera, A. P., Daubaris, P., Bravo, I., Gallego, J., Morabito, R., Widmer, J., & Champati, J. P. (2025) · IEEE Internet of Things Journal · doi:10.1109/jiot.2025.3583477 · Cited by 16
“Thus, significant efforts have been directed toward designing and developing compact DL models tailored explicitly for deployment on end devices [2], [3], [4].”
Mishra, V., Saxena, D., Gupta, K., Patni, S., & Singh, A. K. (2025) · Scientific Reports · doi:10.1038/s41598-025-17937-8 · Cited by 13
“Energy-efficient hardware and renewable energy sources are recommended to train these LLMs 16,17 .”
InfluentialDeployment-Aware Compression of Large Language Models for Edge Intelligence: A System-Level Survey
Dritsas, E., & Trigka, M. (2026) · IEEE Transactions on Artificial Intelligence · doi:10.1109/tai.2026.3656155 · Cited by 17
“Early works, such as [110], offer foundational overviews of quantization and acceleration but are rooted in pretransformer architecture.”
A Low-Rank CNN Architecture for Real-Time Semantic Segmentation in Visual SLAM Applications
Falaschetti, L., Manoni, L., & Turchetti, C. (2022) · IEEE Open Journal of Circuits and Systems · doi:10.1109/ojcas.2022.3174632 · Cited by 11
Pang, B., Liu, S., Wang, H., Guo, B., Wang, Y., Wang, H., Sheng, Z., Wang, Z., & Yu, Z. (2023) · ACM Transactions on Sensor Networks · doi:10.1145/3630098 · Cited by 11
“While most of the DNN compression techniques (e.g., pruning [10, 11], decomposition [32, 50], and quantization [58, 64]) come at the cost of accuracy loss and have the limited ability to accelerate DNN inference [6].”
BackgroundDuan, C., Yang, J., He, X., Qi, Y., Wang, Y., Wang, Y., He, Z., Yan, B., Wang, X., Jia, X., et al. (2023) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2023.3330819 · Cited by 11
“Numerous effective methods have been proposed to introduce sparsity [36], [37], [38].”
MethodRathi, N., & Roy, K. (2024) · IEEE Transactions on Cognitive and Developmental Systems · doi:10.1109/tcds.2024.3396431 · Cited by 14
“Researchers are exploring various alternatives at the hardware (custom chips [4]), algorithms (model compression [5] and transfer learning [6]), and system-level [7] to circumvent the high energy and compute requirements of deep networks.”
Fișne, A., Yurtsever, M. M. E., & Eken, S. (2024) · Cluster Computing · doi:10.1007/s10586-024-04624-y · Cited by 11
“Also using specialized hardware to speed up training and inference, such as Graphics Processing Units (GPUs), Tensor Processing Units (TPUs), or Field Programmable Gate Arrays (FPGAs), resulting in notable speed gains [17, 18].”
Yu, J., Zhang, L., Cheng, D., Huang, W., Wu, H., & Song, A. (2024) · IEEE Transactions on Neural Systems and Rehabilitation Engineering · doi:10.1109/tnsre.2024.3457830 · Cited by 15
“Now there already exist several popular solutions such as tensor decomposition, network pruning, weight quantization, lightweight model design, and knowledge distillation [18], [19].”
BackgroundWiedemann, S., Shivapakash, S., Becking, D., Wiedemann, P., Samek, W., Gerfers, F., & Wiegand, T. (2021) · IEEE Open Journal of Circuits and Systems · doi:10.1109/ojcas.2021.3083332 · Cited by 13
“References [3], [4] give an excellent overview on the landscape of different approaches and techniques studied in this topic.”
BackgroundPruning-Aware Merging for Efficient Multitask Inference
He, X., Gao, D., Zhou, Z., Tong, Y., & Thiele, L. (2021) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3447548.3467271 · Cited by 14
“Pruning a single-task network reduces the computation cost of a neural network while retaining task inference accuracy by suppressing redundancy in the network [4, 24].”
BackgroundRodríguez, I., Campos, C., & Fdez‐Riverola, F. (2021) · Applied Sciences · doi:10.3390/app11199173 · Cited by 13
Zhu, J., Wang, L., & Han, X. G. (2022) · IEEE/ACM International Conference on Automated Software Engineering (ASE) · doi:10.1145/3551349.3556906 · Cited by 18
“) [11, 39]; compared to the big model (often stored in a well-maintained server), attackers thereby have much more opportunities to access to compressed models and attack them(e.”
BackgroundHardware-Software Co-design for Side-Channel Protected Neural Network Inference
Dubey, A., Cammarota, R., Varna, A. L., Kumar, R., & Aysu, A. (2023) · IEEE International Symposium on Hardware Oriented Security and Trust (HOST) · doi:10.1109/host55118.2023.10133716 · Cited by 11
“Interested readers can refer to the survey by Deng et al [34].”
BackgroundSpartan: A Sparsity-Adaptive Framework to Accelerate Deep Neural Network Training on GPUs
Dong, S., Sun, Y., Agostini, N. B., Karimi, E., Lowell, D., Zhou, J. L., Cano, J., Abellán, J. L., & Kaeli, D. R. (2021) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2021.3067825 · Cited by 11
“For inference, the major source of sparsity occurs after applying weight sparsification and quantization [14].”
MethodWinograd convolution: a perspective from fault tolerance
Xue, X., Huang, H., Liu, C., Luo, T., Zhang, L., & Wang, Y. (2022) · Proceedings of the 59th ACM/IEEE Design Automation Conference · doi:10.1145/3489517.3530531 · Cited by 10
“It attracts worldwide attentions and substantial research efforts have been devoted to improving DNNs from the perspective of precision, performance, energy efficiency, and so on [6].”
BackgroundMethodHermes: Memory-Efficient Pipeline Inference for Large Models on Edge Devices
Han, X., Cai, Z., Zhang, Y., Fan, C., Liu, J., Ma, R., & Buyya, R. (2024) · IEEE International Conference on Computer Design: VLSI in Computers and Processors (ICCD) · doi:10.1109/iccd63220.2024.00076 · Cited by 13
Yang, F., Li, N., Wang, L., Jiang, P., Miao, X., & Wang, X. (2025) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2024.3521394 · Cited by 11
“At the circuit level, a bit-reconfigurable ADC is applied within the systolic array to reduce ADC power consumption based on layer-wise bit-width quantization in neural networks [19], [20], [21].”
A survey on dynamic neural networks: From computer vision to multi-modal sensor fusion
Montello, F., Güldenring, R., Scardapane, S., & Nalpantidis, L. (2026) · Image and Vision Computing · doi:10.1016/j.imavis.2026.105980 · Cited by 12
Deep Learning on RISC-V Platforms at the Edge: A Perspective on the Hardware and Software Support
Agosta, G., Galimberti, A., & Zoni, D. (2025) · ACM Computing Surveys · doi:10.1145/3772277 · Cited by 15
“…review compression and optimization techniques for deep neural networks (DNNs) [33], with emphasis on their suitability for hardware acceleration [42, 175], and provide taxonomies of hardware-software (HW/SW) co-design methods and accelerator-oriented optimizations for efficient edge inference…”
Modeling macroscopic brain dynamics with brain-inspired computing architecture
Zheng, Z., Wei, J., Xu, Y., Li, C., Lu, T., Guo, Q.-L., Ji, X., Guo, H., Wang, G., & Deng, L. (2025) · Nature Communications · doi:10.1038/s41467-025-64470-3 · Cited by 12
Sparse CapsNet with explicit regularizer
Shi, R., Niu, L., & Zhou, R. (2021) · Pattern Recognition · doi:10.1016/j.patcog.2021.108486 · Cited by 10
Learning to Design Accurate Deep Learning Accelerators with Inaccurate Multipliers
Jain, P., Huda, S., Maas, M., Gonzalez, J. E., Stoical, I., & Mirhoseini, A. (2022) · Design, Automation, and Test in Europe · doi:10.23919/date54114.2022.9774607 · Cited by 14
“However, low-bit quantization degrades accuracy [2].”
PDAS: Improving network pruning based on Progressive Differentiable Architecture Search for DNNs
Jiang, W., Chen, Y., Wen, S., Zheng, L., & Jin, H. (2023) · Future Generation Computer Systems · doi:10.1016/j.future.2023.04.011 · Cited by 12
Zhang, T., Zheng, J., & Zou, Y. (2023) · Measurement · doi:10.1016/j.measurement.2023.113118 · Cited by 11
SPFQ: a stochastic algorithm and its error analysis for neural network quantization
Zhang, J., & Saab, R. (2026) · Information and Inference A Journal of the IMA · doi:10.1093/imaiai/iaag011 · Cited by 11
Yao, B., Liu, L., Peng, Y., & Peng, X. (2023) · IEEE Transactions on Instrumentation and Measurement · doi:10.1109/tim.2023.3341126 · Cited by 12
“In early studies, pruning and quantization are essential techniques for reducing the spatial and temporal complexity of NN models [11].”
MethodFiltering Empty Video Frames for Efficient Real-Time Object Detection
Liu, Y., & Kang, K. (2024) · Sensors · doi:10.3390/s24103025 · Cited by 10
“However, model compression [47] usually decreases the inference accuracy.”
BackgroundData-Free Quantization via Pseudo-label Filtering
Fan, C., Wang, Z., Guo, D., & Wang, M. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.00534 · Cited by 17
“Deep Neural Networks (DNNs) [26] have shown tremendous potential in a number of fields, but their high computing costs and storage requirements make the implementation complex, especially on embedded systems and edge devices [9, 36].”
BackgroundReparameterization of Lightweight Transformer for On-Device Speech Emotion Recognition
Zhang, Z., Dong, Z., Xu, W., & Han, J. (2024) · IEEE Internet of Things Journal · doi:10.1109/jiot.2024.3483232 · Cited by 15
Guo, F., Li, S., Yang, H., Dong, C., Chen, Y., & Li, G. (2024) · IEEE Transactions on Industrial Informatics · doi:10.1109/tii.2024.3507183 · Cited by 10
“However, most edge devices, with limited computational and communication resources, have difficulty in supporting the deployment of deep learning models [14].”
Self-citationYuan, H. (2025) · Medicine Advances · doi:10.1002/med4.70001 · Cited by 12
Nousias, S., Pikoulis, E. V., Mavrokefalidis, C., & Lalos, A. S. (2021) · IEEE International Conference on Industrial Cyber-Physical Systems (ICPS) · doi:10.1109/icps49255.2021.9468126 · Cited by 11
A Runtime Reconfigurable Design of Compute-in-Memory based Hardware Accelerator
Lu, A., Peng, X., Luo, Y., Huang, S., & Yu, S. (2021) · Design, Automation, and Test in Europe · doi:10.23919/date51398.2021.9474156 · Cited by 13
GRD: An Ultra-Lightweight SAR Ship Detector Based on Global Relationship Distillation
Zhou, Y., Jiang, X., Chen, L., & Liu, X. (2023) · IEEE Geoscience and Remote Sensing Letters · doi:10.1109/lgrs.2023.3302412 · Cited by 10
An end-to-end sensor scheduling method based on D3QN for underwater passive tracking in UWSNs
Zheng, L., Liu, M., & Zhang, S. (2023) · Journal of Network and Computer Applications · doi:10.1016/j.jnca.2023.103730 · Cited by 10
Metadata
- Authors
- Deng, B. L., Li, G., Han, S., Shi, L., & Xie, Y.
- Affiliations
- Tsinghua University (CN)
- Massachusetts Institute of Technology (US)
- University of California, Santa Barbara (US)
- Venue
- Proceedings of the IEEE
- Published
- 2020-04
- Volume
- 108
- Issue
- 4
- Pages
- 485-532
- ISSN
- 0018-9219 (ISSN-L)
- 1558-2256
- OpenAlex
- W3012561096
- Cited by
- 996
- References
- 334
- Citations by year
- 2026: 131 · 2025: 171 · 2024: 193 · 2023: 174 · 2022: 152 · 2021: 142 · 2020: 31
- FWCI
- 46.83
- Type
- journal article
- Language
- en
- Open access
- closed
- Licence
- Locations
- Proceedings of the IEEE (publishedVersion)
- Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
- Funders
- National Science Foundation (1725447)
- Beijing Academy of Artificial Intelligence (BAAI), Tsinghua University Initiative Scientific Research Program
- Institute for Guo Qiang, Tsinghua University
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
- Keywords
- neural network compression, neural network accelerators, deep neural network, data quantization, network sparsification, tensor decomposition, compact modeling, memory efficiency, embedded systems, hardware architecture
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.