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 601–700 of 1,002
TRACED: A Temporal Graph Neural Networks-based Model for Data Prefetching
Jiang, H., Fu, L., Liu, D., Ren, Z., Chen, Y., & Qiao, L. (2025) · ACM Transactions on Architecture and Code Optimization · doi:10.1145/3747843 · Cited by 3
“Dedicated neural network inference accelerators [36, 37] have also emerged, helping prefetchers achieve real-time online inference.”
Zhan, Z., Zhou, S., Zeng, M., Song, Y., Yu, K., Song, M., Chen, X., Wang, J., Hou, Y., & Zhang, R. (2026) · npj Health Systems · doi:10.1038/s44401-026-00152-6 · Cited by 6
Jin, Y., Liu, X., & Liu, J. (2025) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3746027.3758149 · Cited by 4
“It also involves system-level optimization like model pruning [44] and hardware acceleration [10] so that inference can be done under strict delay constraints.”
An Adaptive Compression Method for Lightweight AI Models of Edge Nodes in Customized Production
Jiang, C. S., Hou, M., & Wang, H. (2026) · Sensors · doi:10.3390/s26020383 · Cited by 3
“Quantization-aware training and post-training quantization have both achieved significant success, enabling efficient inference on microcontrollers and low-power GPUs [19].”
Heterogeneity-Guided Tensor Decomposition for Traffic Flow Imputation
Luo, X., Cheng, S., Yang, Y., Wang, P., Guo, S., & Lu, F. (2026) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2026.3658186 · Cited by 3
“Second, techniques such as model compression, data quantization, and grid sparsification can aid in optimizing the model architecture, enhancing both the e ffi ciency and scalability of the model [55], [56].”
Deep learning-based methods for phenotypic trait extraction in rice panicles
Wang, Z., Li, R., Li, W., Ma, X., Yan, S., Li, M., Hu, B., Tang, M., Zhou, G., Wang, J., et al. (2026) · Frontiers in Plant Science · doi:10.3389/fpls.2025.1730366 · Cited by 3
“Future work could explore model compression techniques, such as pruning and quantization (Deng et al., 2020), to create a lightweight version of OPG-YOLOv8 without a signi fi cant loss of accuracy, thereby broadening its applicability.”
Maisto, V., & Cilardo, A. (2026) · ACM Transactions on Architecture and Code Optimization · doi:10.1145/3799238 · Cited by 3
“This can be achieved either by designing lighter network architectures with more efficient blocks and structures, e.g., MobileNet [23], or compressing models, i.e., model compression [17], either after training (post-training quantization) or during such phase (quantization-aware training).”
Compressing Multi-Task Model for Autonomous Driving via Pruning and Knowledge Distillation
Wang, J., Wu, Q. M. J., Zhang, N., Suto, K., & Zhong, L. (2025) · IEEE Annual Congress on Artificial Intelligence of Things (AIoT) · doi:10.1109/aiot66900.2025.00028 · Cited by 3
“With the number of parameters and tasks growing, the requirements for memory and processing power are increasing [6].”
Khan, F. M. A., Hallaq, M., Abou-Zeid, H., Erak, O., Waqar, O., Hassan, S. A., Alhussein, O., & Hossain, E. (2026) · IEEE Communications Surveys & Tutorials · doi:10.1109/comst.2026.3682638 · Cited by 8
“For instance, [11] provides a comprehensive overview of model compression and hardware acceleration techniques, including pruning, quantization, and tensor decomposition, with a focus on algorithm-hardware co-design.”
Song, F. (2021) · Computational Intelligence and Neuroscience · doi:10.1155/2021/9004797 · Cited by 3
“By adding feature information, the MVCNN perspective pooling was extended to multilevel pooling, and the original framework was improved, which achieved good results in practical application [14].”
MethodRetractedA 1.6-mW Sparse Deep Learning Accelerator for Speech Separation
Yang, C.-C., & Chang, T. (2023) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2023.3235760 · Cited by 4
“Many model compression techniques [12]–[20] have been presented for deep learning models to reduce memory size and computational complexity.”
BackgroundMethodDeep Reinforcement Learning Based Multi-Task Automated Channel Pruning for DNNs
Ma, X., & Fang, W. W. (2023) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn54540.2023.10191092 · Cited by 3
“Model compression, which aims to reduce the model size and accelerate the computation at inference, has thus become an active research topic in recent years [2], [5].”
BackgroundMethodTetriX: Flexible Architecture and Optimal Mapping for Tensorized Neural Network Processing
Jie-Fang, Z., Lu, C., & Zhang, Z. (2024) · IEEE Transactions on Computers · doi:10.1109/tc.2024.3365936 · Cited by 3
“To keep improving the accuracy, DNN models grow larger every year, at a rate of 1.5 to 2 × increase in model size and model complexity every year [1], [2].”
MethodInfluentialA Hardware Accelerator for Sparse Computing Based on NVDLA
Chen, Y., Ma, D., & Mao, J. (2024) · China Semiconductor Technology International Conference (CSTIC) · doi:10.1109/cstic61820.2024.10531990 · Cited by 3
“Sparsity, referring to the presence of a significant proportion of zero values in data structures like matrices, has been identified as an essential property that can be leveraged to optimize computations [1].”
BackgroundDTS: dynamic training slimming with feature sparsity for efficient convolutional neural network
Yin, J., Wang, W., Guo, Z., & Ji, Y. (2024) · Journal of Real-Time Image Processing · doi:10.1007/s11554-024-01511-y · Cited by 3
Zou, Y., & Yang, J. (2024) · Industrial Robot the international journal of robotics research and application · doi:10.1108/ir-05-2024-0233 · Cited by 3
F2Zip: Finetuning-Free Model Compression for Scenario-Adaptive Embedded Vision
Luo, P., Hou, J., Yuan, M., Wu, G., Yao, Y., & Li, X. (2024) · ACM Conference on Embedded Networked Sensor Systems (SenSys) · doi:10.1145/3666025.3699319 · Cited by 3
Pelisco, P. D. M., Gonzales, D. R., Monjardin, C. E. F., & de Jesus, K. L. M. (2024) · International Conference on Machine Learning and Machine Intelligence (MLMI) · doi:10.1145/3696271.3696309 · Cited by 3
Neural Evolutionary Architecture Search for Compact Printed Analog Neuromorphic Circuits
Zhao, H., Pal, P., Hefenbrock, M., Wang, Y., Beigl, M., & Tahoori, M. B. (2024) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2024.3524357 · Cited by 3
“Neural network compression [34], including NAS [35], parameter quantization [36], and network pruning [28], may also aid in compact designs.”
Li, W., Li, Z., & Qiao, J. (2025) · Expert Systems with Applications · doi:10.1016/j.eswa.2025.129007 · Cited by 3
Mixed-Precision Quantization: Make the Best Use of Bits Where They Matter Most
Fang, Y., Chen, L., Chen, Y., Wang, W., & You, C. (2025) · IEEE Journal of Selected Topics in Signal Processing · doi:10.1109/jstsp.2025.3589745 · Cited by 3
“In addition to signal processing and communication systems, low-precision quantization has enabled neural network (NN) compression and acceleration [10].”
Wang, Q., Tamukoh, H., & Morie, T. (2025) · Frontiers in Neuroscience · doi:10.3389/fnins.2025.1656892 · Cited by 3
“The sparsity, which we refer to as the ratio of zero weights to total weights, is typically 20%– 50% (Sze et al., 2017; Deng et al., 2020).”
Hardware accelerator for training with integer backpropagation and probabilistic weight update
Park, H., & Kim, S. (2020) · Advances in computers · doi:10.1016/bs.adcom.2020.11.006 · Cited by 3
Simplified Hardware Implementation of Memoryless Dot Product for Neural Network Inference
Kouretas, I., & Paliouras, V. (2021) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas51556.2021.9401625 · Cited by 3
Multimedia Data Analysis With Edge Computing
Chen, S. (2021) · IEEE Multimedia · doi:10.1109/mmul.2021.3124292 · Cited by 3
Huang, K., Chen, S., Li, B., Claesen, L., Yao, H., Chen, J., Jiang, X., Liu, Z., & Xiong, D. (2022) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2022.102403 · Cited by 3
Jung, S., Lee, J., Noh, H., Yoon, J., & Kung, J. (2022) · IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS) · doi:10.1109/aicas54282.2022.9869905 · Cited by 3
“The quantization of activations and/or weight parameters in convolutional neural networks (CNNs) is one of the most effective methods to reduce the memory requirement and computational complexity [1].”
MethodHardware Implementation of Stochastic Computing-based Morphological Neural Systems
Rosselló, J. L., Font-Rosselló, J., Frasser, C. F., Morán, A., Skibinsky-Gitlin, E. S., Canals, V., & Roca, M. (2022) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas48785.2022.9937549 · Cited by 3
“That’s why several ANN compression techniques have been developed in the last years to tackle this challenge and obtain efficient neural network accelerators [2].”
BackgroundGradient Estimation for Ultra Low Precision POT and Additive POT Quantization
Tesfai, H., Saleh, H. H., Al‐Qutayri, M., Mohammad, B. S., & Stouraitis, T. (2023) · IEEE Access · doi:10.1109/access.2023.3286299 · Cited by 3
“Non-uniform quantization usually tends to match important data points from the distribution of weights [5].”
BackgroundOn Hardware-Aware Design and Optimization of Edge Intelligence
Huai, S., Kong, H., Luo, X., Liu, D., Subramaniam, R., Makaya, C., Lin, Q., & Liu, W. (2023) · IEEE Design and Test · doi:10.1109/mdat.2023.3307558 · Cited by 3
“The second reason to have a hardware model is that some literature reveals that hardware architectures have considerable impact on a model’s performance [5].”
Xu, M., Lam, C., Liang, Y., Ng, B. K., & Im, S. (2023) · doi:10.1109/icsip57908.2023.10270968 · Cited by 3
“For tensor decomposition, there are mainly six approaches: 1) Tucker and classical prolongation (CP); 2) higher-order singular value decomposition (HOSVD); 3) tensorizing and curse of dimensionality; 4) hierarchical tucker (HT) and dimension tree; 5) tensor train (TT); 6) tensor chain (TC) [11].”
MethodMachine Learning Hardware Design for Efficiency, Flexibility, and Scalability [Feature]
Jie-Fang, Z., & Zhang, Z. (2023) · IEEE Circuits and Systems Magazine · doi:10.1109/mcas.2023.3302390 · Cited by 3
“For the model weights, network pruning and other sparsification techniques can be used to zero out a large number of weights in a model with only a small inference accuracy drop [24], [25], [26], [27], [28], [29].”
BackgroundMethodPSQ: An Automatic Search Framework for Data-Free Quantization on PIM-based Architecture
Liu, F., Yang, N., & Jiang, L. (2023) · IEEE International Conference on Computer Design: VLSI in Computers and Processors (ICCD) · doi:10.1109/iccd58817.2023.00084 · Cited by 3
“However, the increasing model size and computation of DNNs hinder their deployment on edge devices [1].”
BackgroundMethodInfluentialPau, D. P., Pisani, A., & Candelieri, A. (2024) · Algorithms · doi:10.3390/a17010022 · Cited by 3
Compressing Neural Networks with Two-Layer Decoupling
Jonghe, J. D., Usevich, K., Dreesen, P., & Ishteva, M. K. (2023) · IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing · doi:10.1109/camsap58249.2023.10403509 · Cited by 4
“Consequently, significant research efforts have focused on compressing and accelerating neural networks such as low-rank factorization methods, filter and weight pruning and knowledge distillation [2], [5], [9], [14].”
BackgroundZhong, Y., Lan, K., Fong, S. J., & Wong, D. (2023) · International Conference on Digital Data Processing (DDP) · doi:10.1109/ddp60485.2023.00015 · Cited by 3
Wan, J., Wang, W., & Zhang, Z. (2024) · Geophysical Journal International · doi:10.1093/gji/ggae235 · Cited by 3
Deep Neural Network Pruning with Progressive Regularizer
Zhou, Y., Zhao, H., Hefenbrock, M., Li, S., Huang, Y., & Beigl, M. (2024) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn60899.2024.10650997 · Cited by 6
“However, for large DNNs, the enormous number of parameters leads to expensive computation and memory demands [5].”
BackgroundHarnessing Orthogonality to Train Low-Rank Neural Networks
Coquelin, D., Flügel, K., Weiel, M., Kiefer, N., Debus, C., Streit, A., & Götz, M. (2024) · Frontiers in artificial intelligence and applications · doi:10.3233/faia240729 · Cited by 5
“Asmost tensors can be expressed as low-rank approximations, these methods are often used to compress neural networks [7].”
MethodOptimizing Deep Learning Efficiency through Algorithm-Hardware Co-design
Santoso, J. T., Wibowo, M. C., & Raharjo, B. (2024) · Journal of Advances in Information Technology · doi:10.12720/jait.15.10.1163-1173 · Cited by 3
“Techniques such as pruning, quantization, and specialized hardware accelerators are crucial in addressing these challenges [44, 45].”
BackgroundEfficiency Analysis of Kolmogorov-Arnold Networks for Visual Data Processing
Hollósi, J. (2024) · SMTS · doi:10.3390/engproc2024079068 · Cited by 3
Adaptive knowledge transfer for data-free low-bit quantization via tiered collaborative learning
Yang, Z., Lin, T., Li, C., Qian, B., Yang, X., & Wei, X. (2025) · Neurocomputing · doi:10.1016/j.neucom.2025.130097 · Cited by 3
Mixed-Precision is All You Need for Efficient Document Image Classification
Shinde, T. S., & Bhardwaj, S. (2025) · IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW) · doi:10.1109/wacvw65960.2025.00142 · Cited by 3
“However, traditional uniform quantization methods often fail to consider the heterogeneous sensitivity of layers to precision reduction, leading to significant accuracy degradation [7, 19].”
Folino, F., Folino, G., Guarascio, M., & Pontieri, L. (2025) · Soft Computing · doi:10.1007/s00500-025-10646-4 · Cited by 3
“This limitation is especially pronounced in Green AI contexts (Adadi 2021; Srivastava 2022; Deng et al. 2020), where restrictions on data access and processing operations further constrain the amount of data that can be utilized for model training.”
Liu, Y., Xu, J., Wang, S., Li, Y., Ji, L., Xie, D. H., & Zhou, J. (2025) · Biosensors · doi:10.3390/bios15050327 · Cited by 3
“…varied imaging scenarios and, thus, enhance model robustness [108], adopting active learning and self-supervised strategies to significantly reduce dependency on manual annotations [109,110], and leveraging model compression [111] methods to facilitate fast, offline inference on mobile devices.”
BDC-Occ: Binarized Deep Convolution Unit for Binarized Occupancy Network
Zhang, Z., Ling, P., Xu, Z., Yang, W., Liao, Q., & Xue, J. (2025) · IEEE Transactions on Circuits and Systems for Video Technology · doi:10.1109/tcsvt.2025.3636155 · Cited by 3
“Therefore, it is necessary to introduce model compression techniques [20] to reduce the computational complexity and parameter count of CNN-based occupancy networks.”
Permutation-Invariant Representation of Neural Networks with Neuron Embeddings
Zhou, R., Muise, C. J., & Hu, T. (2022) · Lecture notes in computer science · doi:10.1007/978-3-031-02056-8_19 · Cited by 4
“This reduces the number of parameters from mn to r ( m + n ) [6].”
BackgroundSMOF: Squeezing More Out of Filters Yields Hardware-Friendly CNN Pruning
Liu, Y., Guan, B., Li, W., Xu, Q., & Quan, S. (2022) · Lecture notes in computer science · doi:10.1007/978-3-031-20497-5_20 · Cited by 3
BWCP: Probabilistic Learning-to-Prune Channels for ConvNets via Batch WhiteningPreprint
Shao, W., Yu, H., Zhang, Z., Xu, H., Li, Z., & Luo, P. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2105.06423 · Cited by 4
“However, despite their great success, deep CNN models often have massive demand on storage, memory bandwidth, and computational power [7], making them difficult to be plugged onto resource-limited platforms, such as portable and mobile devices [8, 7].”
BackgroundPruning-Aware Merging for Efficient Multitask InferencePreprint
He, X., Gao, D., Zhou, Z., Tong, Y., & Thiele, L. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1905.09676 · Cited by 4
“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].”
BackgroundYang, Y., Truong, N. D., Eshraghian, J. K., Nikpour, A., & Kavehei, O. (2021) · medRxiv · doi:10.1101/2021.09.30.21264287 · Cited by 5
PTQ-SL: Exploring the Sub-layerwise Post-training QuantizationPreprint
Yuan, Z., Chen, Y., Xue, C., Zhang, C., Wang, Q., & Sun, G. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2110.07809 · Cited by 4
“Recently, we notice that the processing-in-memory-based hardware accelerators have developed rapidly for network inference (Mittal 2019; Deng et al. 2020).”
BackgroundTiny Reinforcement Learning for Quadruped Locomotion using Decision TransformersPreprint
Akgün, O. E., Cuevas, N., Farias, M. R. T., & Garces, D. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2402.13201 · Cited by 3
“This process can be facilitated by constructing better simulated environments (Xie et al., 2020) (Tan et al., 2018), and including real-world data sources for evaluation (De-sai et al., 2020) (Hwangbo et al., 2019).”
BackgroundTask-Specific Zero-Shot Quantization-Aware Training for Object Detection
Li, C., Chen, X., Wang, J., Zhao, K., & Chen, J. (2025) · 2025 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51701.2025.02123 · Cited by 2
“As the demand for deploying deep neural networks on resource-constrained devices continues to grow, quantization has emerged as a critical technique to reduce network size and computational complexity while maintaining performance [7, 11, 20, 52].”
FPGA-QHAR: Throughput-Optimized for Quantized Two-Stream Human Action Recognition on the Edge
Alhussain, A., & Lin, M. (2023) · IEEE International Conference on Smart Communities: Improving Quality of Life using AI, Robotics and IoT (HONET) · doi:10.1109/honet59747.2023.10374852 · Cited by 2
“Numerous researchers aimed to apply compression methods to reduce HAR models' computational complexity and memory footprint, making them more suitable for edge deployment [8], [9].”
MethodArch-Net: Model conversion and quantization for architecture agnostic model deployment
Fang, S., Xu, W., Feng, Z., Yuan, S. M., Wang, Y., Yang, Y., Ding, W., & Zhou, S. (2025) · Neural Networks · doi:10.1016/j.neunet.2025.107384 · Cited by 2
Enhanced convolutional neural network accelerators with memory optimization for routing applications
Nallabelli, S. P., & Sampath, S. (2025) · PLoS ONE · doi:10.1371/journal.pone.0320813 · Cited by 2
“[13].”
Accelerating Edge Intelligence: Challenges and Future Directions in Hardware-Driven Machine Learning
Batzolis, E., & Karampatzakis, D. (2024) · Pan-Hellenic Conference on Progress in Computing and Informatics (PCI) · doi:10.1145/3716554.3716579 · Cited by 2
“Important performance metrics of embedded systems that are of importance to machine learning are: [6][4]: • Model Size: This metric indicates the amount of memory needed to reside the model, the most important it especially for getting models on embedded systems with little storage space.”
Efficient SRAM-PIM Co-Design by Joint Exploration of Value-Level and Bit-Level Sparsity
Duan, C., Yang, J., Wang, Y., Wang, Y., Qi, Y., He, X., Yan, B., Wang, X., Jia, X., & Zhao, W. S. (2025) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2025.3580335 · Cited by 4
“For example, separable convolutions ( sp-conv ) decompose a standard convolution ( std-conv ) into depthwise ( dw-conv ) and pointwise ( pw-conv ) convolutions, significantly reducing computational cost while maintaining representational power [35].”
Enhancing knowledge distillation via genetic recombination
Cao, Y., Zhou, C., Liu, M., Luo, W., & Luo, X. (2025) · Applied Soft Computing · doi:10.1016/j.asoc.2025.113414 · Cited by 2
Adaptive Quantization of Deep Neural Networks via Layer Importance Estimation
Shinde, T. S., & Naik, S. T. (2025) · Communications in computer and information science · doi:10.1007/978-3-031-93709-5_16 · Cited by 9
A Time-Frequency Joint Wideband CSI Prediction Model and Its Validation Using 5G Commercial Networks
Zhou, T. Y., Qi, Y., Xiang, Z., Liu, L., Ai, B., & Jing, X. (2025) · IEEE Transactions on Vehicular Technology · doi:10.1109/tvt.2025.3606019 · Cited by 2
“In order to further reduce the model complexity, there are four kinds of methods used for model compression including model pruning, tensor decomposition, data quantization and network sparsifica-tion [32].”
Overview of Memory-Efficient Architectures for Deep Learning in Real-Time Systems
Demir, B., Domazet, E., & Mechkaroska, D. (2025) · International Conference on Electronics, Engineering Physics and Earth Science (EEPES) · doi:10.3390/engproc2025104077 · Cited by 5
SSpMM: Efficiently S calable SpMM Kernels Across Multiple Generations of Tensor Cores
Xue, Z., Wen, M., Yang, J., Tang, M., Luo, Z., Feng, J., Shi, Y., Chen, Z., Shen, J., & Langguth, J. (2025) · IEEE Transactions on Parallel and Distributed Systems · doi:10.1109/tpds.2025.3616981 · Cited by 3
“These algorithms aim to reduce the model based on different objectives [40], [41], [42], [43], [44], resulting in multi-dimension sparse patterns (i.e., 0-”
Hou, H., Zhang, J., Huang, Z., Zhang, M., & Ye, X. (2025) · The Computer Journal · doi:10.1093/comjnl/bxaf129 · Cited by 2
Li, Y., Zhang, Y., Wang, X., Zhao, Z., Xing, G., Liu, L., Dong, F., Guo, Z., Gao, X., & Xing, G. (2025) · Journal of Applied Physics · doi:10.1063/5.0293805 · Cited by 2
Zhang, J., Che, X., Peng, S., Chen, G., Ma, Q., Fan, B., & Hu, J. (2025) · Applied Soft Computing · doi:10.1016/j.asoc.2025.114273 · Cited by 2
MTAKD: multi-teacher agreement knowledge distillation for edge AI skin disease diagnosis
Winata, A., Andryani, N. A. C., Gunawan, A. A. S., Gaol, F. L., & Matsuo, T. (2025) · Scientific Reports · doi:10.1038/s41598-025-27038-1 · Cited by 3
Hardware-Aware Optimization of Large Language Models for Enhanced Throughput on Arm CPUs
Zhu, X., Chen, L., Zhang, C., Yang, P., Yang, T., Yu, G., Pan, E., Wu, X., Li, B., Mao, W., et al. (2025) · IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS) · doi:10.1109/aicas64808.2025.11173152 · Cited by 2
Qiao, X., Yang, Y., Song, J., Wei, R., Tang, X., Li, M., Wang, R. S., & Wang, Y. (2026) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2026.3651775 · Cited by 2
“For example, 88.1% parameters can be reduced by DSC in a convolution with 128 input channels and 128 output channels [18].”
InfluentialDistilling knowledge for low-energy AIoT
di Torrepadula, F. R., Maisto, V., Cilardo, A., & Mazzocca, N. (2026) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2026.103692 · Cited by 2
YOLO-DSNet for Small Target Detection
Xu, H., He, H., Zhi, Q., Yang, Z., & Han, B. (2026) · Applied Sciences · doi:10.3390/app16031493 · Cited by 2
Awan, S. A., Abbasi, M. A., Burdi, A., Panhwar, A. O., & Sathio, A. A. (2026) · VFAST Transactions on Software Engineering · doi:10.21015/vtse.v14i1.2286 · Cited by 2
Liu, Y., Wang, Z.-F., Wang, Y., Fan, Z., & Liu, Y.-L. (2026) · Tunnelling and Underground Space Technology · doi:10.1016/j.tust.2026.107563 · Cited by 2
Hardware Architectures and Optimization Techniques for Convolutional Neural Network Accelerators
Nehete, H., Verma, G., Monga, A., Shukla, A. K., Yadav, S., & Kaushik, B. K. (2026) · Wiley eBooks · doi:10.1002/9781394336203.ch5 · Cited by 2
Zhang, Y., Lu, W., Weng, L., Lin, H., & Xia, M. (2026) · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · doi:10.1109/jstars.2026.3689534 · Cited by 5
“The initial layer quickly captures high-level snow features (e.g., snow/nonsnow boundaries), and deeper layers use efficient convolutions with progressive compression to balance performance and cost [55].”
Jiang, L., Wu, Y., Wang, Y., Long, J., Chen, Y., Zhou, W., Wang, K., & Yu, J. (2026) · Journal of Lightwave Technology · doi:10.1109/jlt.2026.3691989 · Cited by 2
“Despite the superior performance achieved by NN-based techniques in various impairment mitigation tasks, their practical adoption in real-time coherent re-ceiversalsodependscriticallyonimplementationcost,including computationalcomplexity,memoryfootprint,powerconsump-tion,andlatency[36].”
A Proximal Iteratively Reweighted Approach for Efficient Network Sparsification
Wang, H., Yang, X., Shi, Y., & Lin, J. (2020) · IEEE Transactions on Computers · doi:10.1109/tc.2020.3044142 · Cited by 3
Learning Lightweight Neural Networks via Channel-Split Recurrent Convolution
Wu, G. G., Zhang, X., Zhang, Z., Li, Y., Zhou, X., Brinton, C. G., & Liu, Z. (2023) · IEEE Winter Conference on Applications of Computer Vision · doi:10.1109/wacv56688.2023.00385 · Cited by 4
“Nice surveys on this topic are in [9, 46].”
BackgroundPre-RTL DNN Hardware Evaluator With Fused Layer Support
Yang, C.-C., & Chang, T. (2021) · International SoC Design Conference · doi:10.1109/isocc53507.2021.9614027 · Cited by 2
“Many deep learning accelerators (DLAs) [1]-[3] have been proposed to meet high performance needs with large amount of processing elements and different dataflows.”
BackgroundForming-free titanium oxide neuromorphic crossbar array for robotics and AI systems
Игоревич, А. В., Vakulov, Z., Fedotov, A., Tominov, R. V., Polupanov, N. V., & Smirnov, V. A. (2022) · Scientific School Dynamics of Complex Networks and their Applications (DCNA) · doi:10.1109/dcna56428.2022.9923309 · Cited by 2
“In some cases, AI accelerator circuits are already used, containing arrays of transistors that optimize the process of big data neural network processing, but in most cases, the problem of long-term data exchange between the memory and the AI accelerator of the computing device remains [12-15].”
BackgroundCombining Non-Data-Adaptive Transforms for OCT Image Denoising by Iterative Basis Pursuit
Razavi, R., Rabbani, H., & Plonka, G. (2022) · Proceedings - International Conference on Image Processing · doi:10.1109/icip46576.2022.9897319 · Cited by 2
Wang, H., Du, J., Zhou, H., Lu, H., & Cao, Y. (2022) · Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · doi:10.23919/apsipaasc55919.2022.9980299 · Cited by 2
“Network pruning includes structured pruning and unstructured pruning [16, 17].”
Bocharov, N., Paramonov, N., Slavin, O., & Suminov, K. (2022) · Proceedings of the Institute for System Programming of RAS · doi:10.15514/ispras-2022-34(6)-6 · Cited by 2
Li, C., Xu, M., Yang, J., Yao, Y., Wang, S., Gao, X., Shao, S., Chen, X., Zheng, X., & Liu, Y. (2023) · Journal of Industrial and Management Optimization · doi:10.3934/jimo.2023027 · Cited by 3
TinyML Techniques for running Machine Learning models on Edge Devices
Mukherjee, A., Ukil, A., Dey, S., & Kulkarni, G. (2022) · International Conference on AI-ML Systems (AIMLSystems) · doi:10.1145/3564121.3564812 · Cited by 3
“Hence, model compression is of utmost importance to put such models to use with minimum penalty in tiny edge devices [3, 4, 6, 10, 14, 17, 21].”
BackgroundLin, J., Ye, Z., & Wang, J. (2023) · Asia-Europe Conference on Electronics, Data Processing and Informatics (ACEDPI) · doi:10.1109/acedpi58926.2023.00058 · Cited by 2
Revisit and Benchmarking of Automated Quantization Toward Fair Comparison
Wei, Z., Zhang, X., Ji, Z., Li, J., & Wei, J. (2023) · IEEE Transactions on Computers · doi:10.1109/tc.2023.3315836 · Cited by 3
“Hardware-aware automated quantization has emerged as a novel co-design paradigm where the automated quantization algorithm guides efficient hardware design, and the hardware design provides the feedback for the high-level algorithm [14].”
MethodWhen Side-Channel Attacks Break the Black-Box Property of Embedded Artificial Intelligence
Coqueret, B., Carbone, M., Sentieys, O., & Zaid, G. (2023) · ACM Workshop on Artificial Intelligence and Security (AISec) · doi:10.1145/3605764.3623903 · Cited by 5
Pruning Deep Neural Network Models via Minimax Concave Penalty Regression
Liu, X., Zhou, L. Y., & Luo, Y. (2024) · Applied Sciences · doi:10.3390/app14093669 · Cited by 2
On energy complexity of fully-connected layers
Šíma, J., Cabessa, J., & Vidnerová, P. (2024) · Neural Networks · doi:10.1016/j.neunet.2024.106419 · Cited by 3
“…enormous amount of energy can be saved at the cost of only a small loss in accuracy by using approximate computing methods (Armeniakos et al., 2023; Deng et al., 2020; Li et al., 2023; Lyu et al., 2023; Mittal, 2016; Tang et al., 2024), e.g. low float precision (Gupta et al., 2015), approximate…”
MethodCost-Efficient SIMD ASIP Architecture for Mobile Touchscreen Controllers
Park, E., Jin, M., & Lee, Y. (2024) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas58744.2024.10557834 · Cited by 2
“Unlike the straightforward single-instruction multiple-data (SIMD) approaches that are normally specialized for general matrix multiplications [8], more TDA-dedicated SIMD instructions are newly defined by considering the available DLP and accessing patterns in the standard TDA [9].”
Bridging Dimensions: Confident Reachability for High-Dimensional Controllers
Geng, Y., Baldauf, J. B., Dutta, S., Huang, C., & Ruchkin, I. (2024) · Lecture notes in computer science · doi:10.1007/978-3-031-71162-6_20 · Cited by 10
Correlation-Aware Joint Pruning-Quantization using Graph Neural Networks
Nor-Azman, M. N. A., Sheikh, U. U., Mohammed, M. S., Sirkunan, J., & Marsono, M. N. (2024) · Proceedings - International Conference on Image Processing · doi:10.1109/icip51287.2024.10647517 · Cited by 2
SearchQ: Search-Based Fine-Grained Quantization for Data-Free Model Compression
Yang, N., Liu, F., Wang, Z., Zhao, J., & Jiang, L. (2024) · IEEE transactions on circuits and systems for artificial intelligence. · doi:10.1109/tcasai.2024.3491941 · Cited by 3
LiDAR点云深度学习模型的压缩和部署加速方法研究现状与展望(特邀)
Luda, 赵. Z., Yihua, 胡. H., Nanxiang, 赵. Z., Fei, 汪. W., Yicheng, 王. W., Xiao, 董. D., Fei, 韩. H., Zhenglei, 豆. D., & Hou, A. (2024) · Laser & Optoelectronics Progress · doi:10.3788/lop241166 · Cited by 3
MCD: Multi-stage Catalytic Distillation for Time Series Forecasting
Ma, R., Zhang, C., Kang, Y., Wang, X., & Qiu, C. (2024) · Lecture notes in computer science · doi:10.1007/978-981-97-5569-1_34 · Cited by 2
Research on Anti-Interference Performance of Spiking Neural Network Under Network Connection Damage
Zhang, Y., Pang, H., Ma, J., Ma, G., Zhang, X., & Man, M. (2025) · Brain Sciences · doi:10.3390/brainsci15030217 · Cited by 2
“It reduces the resource requirements of the network, including memory, computing, and energy consumption [24].”
Berg, O. A. B., Saqib, E., Jantsch, A., Shallari, I., Krug, S., Leal, I. S., & O’Nils, M. (2025) · IEEE Access · doi:10.1109/access.2025.3579107 · Cited by 2
“However, deploying deep learning models on these resource-constrained devices presents challenges due to their limited processing capacity [1], [2].”
Hassan, A., Sjöström, M., Zhang, T., & Egiazarian, K. (2025) · IEEE Transactions on Emerging Topics in Computational Intelligence · doi:10.1109/tetci.2025.3592281 · Cited by 3
“If needed, convolution operations with small kernels (e.g., 3 × 3 ) can be stacked multiple times to approximate the effect of larger kernels (e.g., 5 × 5 or 7 × 7 ), offering greater efficient as each layer incrementally expands the receptive field during features extraction [55].”
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.