README.md

A Comprehensive Survey on AI Integration at the Edge: Techniques, Applications, and Challenges

Xubin Wang¹² · Zhiqing Tang² · Jianxiong Guo²³ · Tianhui Meng² · Chenhao Wang²³ · Tian Wang²³ · Weijia Jia^23*

¹Hong Kong Baptist University · ²Beijing Normal University · ³BNU-HKBU United International College

^*corresponding authors

Table of Contents

• 1. Background Knowledge
  • 1.1. Edge Computing
  • 1.2. Edge AI
    • 1.2.1. Blogs About Edge AI
• 2. Our Survey
  • 2.1. The Taxonomy of the Discussed Topics
  • 2.2. Edge AI Optimization Triad
  • 2.3. The Edge AI Deployment Pipeline
• 3. The Data-Model-System Optimization Triad
  • 3.1. Data Optimization
    • 3.1.1. Data Cleaning
    • 3.1.2. Feature Compression
      • 3.1.2.1. Feature Selection
      • 3.1.2.2. Feature Extraction
    • 3.1.3. Data Augmentation
  • 3.2. Model Optimization
    • 3.2.1. Model Design
      • 3.2.1.1. Compact Architecture Design
      • 3.2.1.2. Neural Architecture Search (NAS)
    • 3.2.2. Model Compression
      • 3.2.2.1. Model Pruning
      • 3.2.2.2. Parameter Sharing
      • 3.2.2.3. Model Quantization
      • 3.2.2.4. Knowledge Distillation
      • 3.2.2.5. Low-rank Factorization
  • 3.3. System Optimization
    • 3.3.1. Software Optimization
    • 3.3.2. Hardware Optimization

1. Background Knowledge

1.1. Edge Computing

Edge computing is a distributed computing paradigm that brings computation and data storage closer to the sources of data. This is expected to improve response times and save bandwidth.

• What is edge computing? Everything you need to know
• Machine Learning at the Edge — μML

1.2. Edge AI

• Edge AI refers to the deployment of artificial intelligence (AI) algorithms and models directly on edge devices, such as mobile phones, Internet of Things (IoT) devices, and other smart sensors.
• By processing data locally on the device rather than relying on cloud-based algorithms, edge AI enables real-time decision-making and reduces the need for data to be transmitted to remote servers. This can lead to reduced latency, improved data privacy and security, and reduced bandwidth requirements.
• Edge AI has become increasingly important as the proliferation of edge devices continues to grow, and the need for intelligent and low-latency decision-making becomes more pressing.

1.2.1. Blogs About Edge AI

• Edge AI – What is it and how does it Work?
• What is Edge AI?
• Edge AI – Driving Next-Gen AI Applications in 2022
• Edge Intelligence: Edge Computing and Machine Learning (2023 Guide)
• What is Edge AI, and how does it work?
• Edge AI 101- What is it, Why is it important, and How to implement Edge AI?
• Edge AI: The Future of Artificial Intelligence
• What is Edge AI? Machine Learning + IoT
• What is edge AI computing?
• 在边缘实现机器学习都需要什么？
• 边缘计算 | 在移动设备上部署深度学习模型的思路与注意点

2. Our Survey (To be released)

2.1 The Taxonomy of the Discussed Topics

2.2 Edge AI Optimization Triad

We introduce a data-model-system optimization triad for edge deployment.

2.3 The Edge AI Deployment Pipeline

An overview of edge deployment. The figure shows a general pipeline from the three aspects of data, model and system. Note that not all steps are necessary in real applications.

3. The Data-Model-System Optimization Triad

3.1. Data Optimization

An overview of data optimization operations. Data cleaning improves data quality by removing errors and inconsistencies in the raw data. Feature compression is used to eliminate irrelevant and redundant features. For scarce data, data augmentation is employed to increase the data size.

3.1.1. Data Cleaning

Title & Basic Information	Affiliation	Code
Active label cleaning for improved dataset quality under resource constraints[J]. Nature communications, 2022.	Microsoft Research Cambridge	Code
Locomotion mode recognition using sensory data with noisy labels: A deep learning approach. IEEE Trans. on Mobile Computing.	Indian Institute of Technology BHU Varanasi	Code
Big data cleaning based on mobile edge computing in industrial sensor-cloud[J]. IEEE Trans. on Industrial Informatics, 2019.	Huaqiao University	--
Federated data cleaning: Collaborative and privacy-preserving data cleaning for edge intelligence[J]. IoTJ, 2020.	Xidian University	--
A data stream cleaning system using edge intelligence for smart city industrial environments[J]. IEEE Trans. on Industrial Informatics, 2021.	Hangzhou Dianzi University	--
Protonn: Compressed and accurate knn for resource-scarce devices[C] ICML, 2017.	Microsoft Research, India	Code
Intelligent data collaboration in heterogeneous-device iot platforms[J]. ACM Trans. on Sensor Networks (TOSN), 2021.	Hangzhou Dianzi University	--

3.1.2. Feature Compression

3.1.2.1. Feature Selection

Title & Basic Information	Affiliation	Code
Accessible melanoma detection using smartphones and mobile image analysis[J]. IEEE Trans. on Multimedia, 2018.	Singapore University of Technology and Design	--
ActID: An efficient framework for activity sensor based user identification[J]. Computers & Security, 2021.	University of Houston-Clear Lake	--
Descriptor Scoring for Feature Selection in Real-Time Visual Slam[C] ICIP, 2020.	Processor Architecture Research Lab, Intel Labs	--
Edge2Analysis: a novel AIoT platform for atrial fibrillation recognition and detection[J]. IEEE Journal of Biomedical and Health Informatics, 2022.	Sun Yat-Sen University	--
Feature selection with limited bit depth mutual information for portable embedded systems[J]. Knowledge-Based Systems, 2020.	CITIC, Universidade da Coruña	--
Seremas: Self-resilient mobile autonomous systems through predictive edge computing[C] SECON, 2021.	University of California, Irvine	--
A covid-19 detection algorithm using deep features and discrete social learning particle swarm optimization for edge computing devices[J]. ACM Trans. on Internet Technology (TOIT), 2021.	Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System	--

3.1.2.2. Feature Extraction

Title & Basic Information	Affiliation	Code
Supervised compression for resource-constrained edge computing systems[C] WACV, 2022.	University of California, Irvine	Code
"Blessing of dimensionality: High-dimensional feature and its efficient compression for face verification." CVPR, 2013.	University of Science and Technology of China	--
Toward intelligent sensing: Intermediate deep feature compression[J]. TIP, 2019.	Nangyang Technological University	--
Selective feature compression for efficient activity recognition inference[C] ICCV, 2021.	Amazon Web Services	--
Video coding for machines: A paradigm of collaborative compression and intelligent analytics[J]. TIP, 2020.	Peking University	--
Communication-computation trade-off in resource-constrained edge inference[J]. IEEE Communications Magazine, 2020.	The Hong Kong Polytechnic University	Code
Edge-based compression and classification for smart healthcare systems: Concept, implementation and evaluation[J]. ESWA, 2019.	Qatar University	--
EFCam: Configuration-adaptive fog-assisted wireless cameras with reinforcement learning[C] SECON, 2021.	Nanyang Technological University	--
Edge computing for smart health: Context-aware approaches, opportunities, and challenges[J]. IEEE Network, 2019.	Qatar University	--
DEEPEYE: A deeply tensor-compressed neural network for video comprehension on terminal devices[J]. TECS, 2020.	Shanghai Jiao Tong University	--
CROWD: crow search and deep learning based feature extractor for classification of Parkinson’s disease[J]. TOIT, 2021.	Taif University	--
"Deep-Learning Based Monitoring Of Fog Layer Dynamics In Wastewater Pumping Stations", Water research 202 (2021): 117482.	Deltares	--
Distributed and efficient object detection via interactions among devices, edge, and cloud[J]. IEEE Trans. on Multimedia, 2019.	Central South University	--

3.1.3. Data Augmentation

Title & Basic Information	Affiliation	Code
An effective litchi detection method based on edge devices in a complex scene[J]. Biosystems Engineering, 2022.	Beihang University	--
Segmentation of drivable road using deep fully convolutional residual network with pyramid pooling[J]. Cognitive Computation, 2018.	Tsinghua University	--
Multiuser physical layer authentication in internet of things with data augmentation[J]. IoTJ, 2019.	University of Electronic Science and Technology of China	--
Data-augmentation-based cellular traffic prediction in edge-computing-enabled smart city[J]. TII, 2020.	University of Electronic Science and Technology of China	--
Towards light-weight and real-time line segment detection[C] AAAI, 2022.	NAVER/LINE Corp.	Code
Intrusion Detection System After Data Augmentation Schemes Based on the VAE and CVAE[J]. IEEE Trans. on Reliability, 2022.	Guangdong Ocean University	--
Magicinput: Training-free multi-lingual finger input system using data augmentation based on mnists[C] ICIP, 2021.	Shanghai Jiao Tong University	--

3.2. Model Optimization

An overview of model optimization operations. Model design involves creating lightweight models through manual and automated techniques, including architecture selection, parameter tuning, and regularization. Model compression involves using various techniques, such as pruning, quantization, and knowledge distillation, to reduce the size of the model and obtain a compact model that requires fewer resources while maintaining high accuracy.

3.2.1. Model Design

3.2.1.1. Compact Architecture Design

Title & Basic Information	Affiliation	Code
Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]. arXiv, 2017.	Google Inc.	Code
Mobilenetv2: Inverted residuals and linear bottlenecks[C] CVPR, 2018.	Google Inc.	Code
Searching for mobilenetv3[C]// ICCV, 2019.	Google Inc.	Code
Rethinking bottleneck structure for efficient mobile network design[C] ECCV, 2020.	National University of Singapore	Code
Mnasnet: Platform-aware neural architecture search for mobile[C] CVPR, 2019.	Google Brain	Code
Shufflenet: An extremely efficient convolutional neural network for mobile devices[C] CVPR, 2018.	Megvii Inc (Face++)	Code
Shufflenet v2: Practical guidelines for efficient cnn architecture design[C] ECCV, 2018.	Megvii Inc (Face++)	Code
Single path one-shot neural architecture search with uniform sampling[C] ECCV, 2020.	Megvii Inc (Face++)	Code
SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size[J]. arXiv, 2016.	DeepScale∗ & UC Berkeley	Code
Squeezenext: Hardware-aware neural network design[C] CVPR Workshops. 2018.	UC Berkeley	Code
Ghostnet: More features from cheap operations[C] CVPR, 2020.	Noah’s Ark Lab, Huawei Technologies	Code
Efficientnet: Rethinking model scaling for convolutional neural networks[C] ICML, 2019.	Google Brain	Code
Efficientnetv2: Smaller models and faster training[C] ICML, 2021.	Google Brain	Code
Efficientdet: Scalable and efficient object detection[C] CVPR, 2020.	Google Brain	Code
Condensenet: An efficient densenet using learned group convolutions[C] CVPR, 2018.	Cornell University	Code
Condensenet v2: Sparse feature reactivation for deep networks[C] CVPR, 2021.	Tsinghua University	Code
Espnet: Efficient spatial pyramid of dilated convolutions for semantic segmentation[C] ECCV, 2018.	University of Washington	Code
Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network[C] CVPR, 2019.	University of Washington	Code
Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search[C] CVPR, 2019.	UC Berkeley	Code
Fbnetv2: Differentiable neural architecture search for spatial and channel dimensions[C] CVPR, 2020.	UC Berkeley	Code
Fbnetv3: Joint architecture-recipe search using predictor pretraining[C] CVPR, 2021.	Facebook Inc. & UC Berkeley	--
Pelee: A real-time object detection system on mobile devices[J]. NeurIPS, 2021.	University of Western Ontario	Code
Going deeper with convolutions[C] CVPR, 2015.	Google Inc.	Code
Batch normalization: Accelerating deep network training by reducing internal covariate shift[C] ICML, 2015.	Google Inc.	Code
Rethinking the inception architecture for computer vision[C]// CVPR, 2016.	Google Inc.	Code
Inception-v4, inception-resnet and the impact of residual connections on learning[C] AAAI, 2017.	Google Inc.	Code
Xception: Deep learning with depthwise separable convolutions[C] CVPR, 2017.	Google, Inc.	Code
Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer[J]. arXiv, 2021.	Apple	Code
Lite transformer with long-short range attention[J]. arXiv, 2020.	Massachusetts Institute of Technology	Code
Coordinate attention for efficient mobile network design[C] CVPR, 2021.	National University of Singapore	Code
ECA-Net: Efficient channel attention for deep convolutional neural networks[C] CVPR, 2020.	Tianjin University	Code
Sa-net: Shuffle attention for deep convolutional neural networks[C] ICASSP, 2021.	Nanjing University	Code
Triplet Attention: Rethinking the Similarity in Transformers[C] KDD, 2021.	Beihang University	Code
Resnest: Split-attention networks[C] CVPR, 2020.	Meta	Code

3.2.1.2. Neural Architecture Search (NAS)

Title & Basic Information	Affiliation	Code
FTT-NAS: Discovering fault-tolerant convolutional neural architecture[J]. TODAES), 2021.	Tsinghua University	Code
An adaptive neural architecture search design for collaborative edge-cloud computing[J]. IEEE Network, 2021.	Nanjing University of Posts and Telecommunications	--
Binarized neural architecture search for efficient object recognition[J]. IJCV, 2021.	Beihang University	--
Multiobjective reinforcement learning-based neural architecture search for efficient portrait parsing[J]. IEEE Trans. on Cybernetics, 2021.	University of Electronic Science and Technology of China	--
Intermittent-aware neural architecture search[J]. ACM Transactions on Embedded Computing Systems (TECS), 2021.	Academia Sinica and National Taiwan University	Code
Hardcore-nas: Hard constrained differentiable neural architecture search[C] ICML, 2021.	Alibaba Group, Tel Aviv, Israel	Code
MemNAS: Memory-efficient neural architecture search with grow-trim learning[C] CVPR, 2020.	Beijing University of Posts and Telecommunications	--
Pvnas: 3D neural architecture search with point-voxel convolution[J]. TPAMI, 2021.	Massachusetts Institute of Technology	--
Toward tailored models on private aiot devices: Federated direct neural architecture search[J]. IoTJ, 2022.	Northeastern University, Qinhuangdao	--
Automatic design of convolutional neural network architectures under resource constraints[J]. TNNLS, 2021.	Sichuan University	--

3.2.2. Model Compression

3.2.2.1. Model Pruning

Title & Basic Information	Affiliation	Code
Supervised compression for resource-constrained edge computing systems[C] WACV, 2022.	University of Pittsburgh	--
Train big, then compress: Rethinking model size for efficient training and inference of transformers[C] ICML, 2020.	UC Berkeley	--
Hrank: Filter pruning using high-rank feature map[C]// CVPR, 2020.	Xiamen University	Code
Clip-q: Deep network compression learning by in-parallel pruning-quantization[C] CVPR, 2018.	Simon Fraser University	--
Sparse: Sparse architecture search for cnns on resource-constrained microcontrollers[J]. NeurIPS, 2019.	Arm ML Research	--
Deepadapter: A collaborative deep learning framework for the mobile web using context-aware network pruning[C] INFOCOM, 2020.	Beijing University of Posts and Telecommunications	--
SCANN: Synthesis of compact and accurate neural networks[J]. IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems, 2021.	Princeton University	--
Directx: Dynamic resource-aware cnn reconfiguration framework for real-time mobile applications[J]. IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems, 2020.	George Mason University	--
Pruning deep reinforcement learning for dual user experience and storage lifetime improvement on mobile devices[J]. IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems, 2020.	City University of Hong Kong	--
SuperSlash: A unified design space exploration and model compression methodology for design of deep learning accelerators with reduced off-chip memory access volume[J]. IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems, 2020.	Information Technology University	--
Penni: Pruned kernel sharing for efficient CNN inference[C] ICML, 2020.	Duke University	Code
Fast operation mode selection for highly efficient iot edge devices[J]. IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems, 2019.	Karlsruhe Institute of Technology	--
Efficient on-chip learning for optical neural networks through power-aware sparse zeroth-order optimization[C] AAAI, 2021.	University of Texas at Austin	--
A Fast Post-Training Pruning Framework for Transformers[C]// NeurIPS	UC Berkeley	Code
Radio frequency fingerprinting on the edge[J]. TMC, 2021.	Northeastern University, Boston	--
Exploring sparsity in image super-resolution for efficient inference[C] CVPR, 2021.	National University of Defense Technology	Code
O3BNN-R: An out-of-order architecture for high-performance and regularized BNN inference[J]. TPDS, 2020.	Boston University	--
Enabling on-device cnn training by self-supervised instance filtering and error map pruning[J]. TCAD, 2020.	University of Pittsburgh	--
Dropnet: Reducing neural network complexity via iterative pruning[C] ICML, 2020.	National University of Singapore	Code
Edgebert: Sentence-level energy optimizations for latency-aware multi-task nlp inference[C] MICRO-54, 2021.	Harvard University	Code
Fusion-catalyzed pruning for optimizing deep learning on intelligent edge devices[J]. TCAD, 2020.	Chinese Academy of Sciences	--
3D CNN acceleration on FPGA using hardware-aware pruning[C] DAC, 2020.	Northeastern University, MA	--
Width & depth pruning for vision transformers[C] AAAI, 2020.	Institute of Computing Technology, Chinese Academy of Sciences	--
Prive-hd: Privacy-preserved hyperdimensional computing[C] DAC, 2020.	UC San Diego	--
NestFL: efficient federated learning through progressive model pruning in heterogeneous edge computing[C] MobiCom, 2022.	Purple Mountain Laboratories, Nanjing	--

3.2.2.2. Parameter Sharing

Title & Basic Information	Affiliation	Code
Deep k-means: Re-training and parameter sharing with harder cluster assignments for compressing deep convolutions[C] ICML, 2018.	Texas A&M University	Code
T-basis: a compact representation for neural networks[C] ICML, 2020.	ETH Zurich	Code
"Soft Weight-Sharing for Neural Network Compression." International Conference on Learning Representations.	University of Amsterdam	Code
ShiftAddNAS: Hardware-inspired search for more accurate and efficient neural networks[C] ICML, 2022.	Rice University	Code
EfficientTDNN: Efficient architecture search for speaker recognition[J]. IEEE/ACM Trans. on Audio, Speech, and Language Processing, 2022.	Tongji University	Code
A generic network compression framework for sequential recommender systems[C] SIGIR, 2020.	University of Science and Technology	Code
Neural architecture search for LF-MMI trained time delay neural networks[J]. IEEE/ACM Trans. on Audio, Speech, and Language Processing, 2022.	The Chinese University of Hong Kong	--
Structured transforms for small-footprint deep learning[J]. NeurIPS, 2015.	Google, New York	--

3.2.2.3. Model Quantization

Title & Basic Information	Affiliation	Code
Fractrain: Fractionally squeezing bit savings both temporally and spatially for efficient dnn training[J]. NeurIPS, 2020.	Rice University	Code
Edgebert: Sentence-level energy optimizations for latency-aware multi-task nlp inference[C] MICRO-54, 2021.	Harvard University	--
Stochastic precision ensemble: self-knowledge distillation for quantized deep neural networks[C] AAAI, 2021.	Seoul National University	--
Q-capsnets: A specialized framework for quantizing capsule networks[C] DAC, 2020.	Technische Universitat Wien (TU Wien)	Code
Fspinn: An optimization framework for memory-efficient and energy-efficient spiking neural networks[J]. TCAD, 2020.	Technische Universität Wien	--
Octo: INT8 Training with Loss-aware Compensation and Backward Quantization for Tiny On-device Learning[C] USENIX Annual Technical Conference. 2021.	Hong Kong Polytechnic University	Code
Hardware-centric automl for mixed-precision quantization[J]. IJCV, 2020.	Massachusetts Institute of Technology	--
An automated quantization framework for high-utilization rram-based pim[J]. TCAD, 2021.	Capital Normal University	--
Exact neural networks from inexact multipliers via fibonacci weight encoding[C] DAC, 2021.	Swiss Federal Institute of Technology Lausanne (EPFL)	--
Integer-arithmetic-only certified robustness for quantized neural networks[C] ICCV, 2021.	University of Southern California	--
Bits-Ensemble: Toward Light-Weight Robust Deep Ensemble by Bits-Sharing[J]. TCAD, 2022.	McGill University	--
Similarity-Aware CNN for Efficient Video Recognition at the Edge[J]. TCAD, 2021.	University of Southampton	--
Data-Free Network Compression via Parametric Non-uniform Mixed Precision Quantization[C] CVPR, 2022.	Huawei Noah's Ark Lab	--

3.2.2.4. Knowledge Distillation

Title & Basic Information	Affiliation	Code
Be your own teacher: Improve the performance of convolutional neural networks via self distillation[C] ICCV, 2019.	Tsinghua University	Code
Dynabert: Dynamic bert with adaptive width and depth[J]. NeurIPS, 2020.	Huawei Noah’s Ark Lab	Code
Scan: A scalable neural networks framework towards compact and efficient models[J]. NeurIPS, 2019.	Tsinghua University	Code
Content-aware gan compression[C] CVPR, 2021.	Princeton University	--
Stochastic precision ensemble: self-knowledge distillation for quantized deep neural networks[C] AAAI, 2021.	Seoul National University	--
Cross-modal knowledge distillation for vision-to-sensor action recognition[C] ICASSP, 2022.	Texas State University	--
Learning efficient and accurate detectors with dynamic knowledge distillation in remote sensing imagery[J]. IEEE Trans. on Geoscience and Remote Sensing, 2021.	Chinese Academy of Sciences	--
On-Device Next-Item Recommendation with Self-Supervised Knowledge Distillation[C] SIGIR, 2022.	The University of Queensland	Code
Personalized edge intelligence via federated self-knowledge distillation[J]. TPDS, 2022.	Huazhong University of Science and Technology	--
Mobilefaceswap: A lightweight framework for video face swapping[C] AAAI, 2022.	Baidu Inc.	--
Dynamically pruning segformer for efficient semantic segmentation[C] ICASSP, 2022.	Amazon Halo Health & Wellness	--
CDFKD-MFS: Collaborative Data-Free Knowledge Distillation via Multi-Level Feature Sharing[J]. IEEE Trans. on Multimedia, 2022.	Beijing Institute of Technology	Code
Learning Efficient Vision Transformers via Fine-Grained Manifold Distillation[J]. NeurIPS, 2022.	Beijing Institute of Technology	Code
Learning Accurate, Speedy, Lightweight CNNs via Instance-Specific Multi-Teacher Knowledge Distillation for Distracted Driver Posture Identification[J]. IEEE Trans. on Intelligent Transportation Systems, 2022.	Hefei Institutes of Physical Science (HFIPS), Chinese Academy of Sciences	--

3.2.2.5. Low-rank Factorization

Title & Basic Information	Affiliation	Code
Learning low-rank deep neural networks via singular vector orthogonality regularization and singular value sparsification[C] CVPR workshops. 2020.	Duke University	--
MicroNet: Towards image recognition with extremely low FLOPs[J]. arXiv, 2020.	UC San Diego	--
Locality Sensitive Hash Aggregated Nonlinear Neighborhood Matrix Factorization for Online Sparse Big Data Analysis[J]. ACM/IMS Transactions on Data Science (TDS), 2022.	Hunan University	--

3.3. System Optimization

An overview of system optimization operations. Software optimization involves developing frameworks for lightweight model training and inference, while hardware optimization focuses on accelerating models using hardware-based approaches to improve computational efficiency on edge devices.

3.3.1. Software Optimization

Title & Basic Information	Affiliation	Code
Hidet: Task-mapping programming paradigm for deep learning tensor programs[C] ASPLOS Conference, 2023.	University of Toronto	Code
SparkNoC: An energy-efficiency FPGA-based accelerator using optimized lightweight CNN for edge computing[J]. Journal of Systems Architecture, 2021.	Shanghai Advanced Research Institute, Chinese Academy of Sciences	--
Re-architecting the on-chip memory sub-system of machine-learning accelerator for embedded devices[C] ICCAD, 2016.	Institute of Computing Technology, Chinese Academy of Sciences	--
A unified optimization approach for cnn model inference on integrated gpus[C] ICPP, 2019.	Amazon Web Services	Code
ACG-engine: An inference accelerator for content generative neural networks[C] ICCAD, 2019.	University of Chinese Academy of Sciences	--
Edgeeye: An edge service framework for real-time intelligent video analytics[C] EDGESYS Conference, 2018.	University of Wisconsin-Madison	--
Haq: Hardware-aware automated quantization with mixed precision[C] CVPR, 2019.	Massachusetts Institute of Technology	--
Source compression with bounded dnn perception loss for iot edge computer vision[C] MobiCom, 2019.	Hewlett Packard Labs	--
A lightweight collaborative deep neural network for the mobile web in edge cloud[J]. TMC, 2020.	Beijing University of Posts and Telecommunications	--
Enabling incremental knowledge transfer for object detection at the edge[C] CVPR Workshops, 2020.	Arizona State university	--
DA3: Dynamic Additive Attention Adaption for Memory-Efficient On-Device Multi-Domain Learning[C] CVPR, 2022.	Arizona State University	--
An efficient GPU-accelerated inference engine for binary neural network on mobile phones[J]. Journal of Systems Architecture, 2021.	Sun Yat-sen University	Code
RAPID-RL: A Reconfigurable Architecture with Preemptive-Exits for Efficient Deep-Reinforcement Learning[C] ICRA, 2022.	Purdue University	--
A variational information bottleneck based method to compress sequential networks for human action recognition[C] WACV, 2021.	Indian Institute of Technology Delhi	--
EdgeDRNN: Recurrent neural network accelerator for edge inference[J]. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2020.	University of Zürich and ETH Zürich	--
Structured pruning of recurrent neural networks through neuron selection[J]. Neural Networks, 2020.	University of Electronic Science and Technology of China	--
Dynamically hierarchy revolution: dirnet for compressing recurrent neural network on mobile devices[J]. arXiv, 2018.	Arizona State University	--
High-throughput cnn inference on embedded arm big. little multicore processors[J]. TCAD, 2019.	National University of Singapore	--
SCA: a secure CNN accelerator for both training and inference[C] DAC, 2020.	University of Pittsburgh	--
NeuLens: spatial-based dynamic acceleration of convolutional neural networks on edge[C] MobiCom, 2022.	New Jersey Institute of Technology	--
Weightless neural networks for efficient edge inference[C] PACT, 2022.	The University of Texas at Austin	Code
O3BNN-R: An out-of-order architecture for high-performance and regularized BNN inference[J]. TPDS, 2020.	--
Blockgnn: Towards efficient gnn acceleration using block-circulant weight matrices[C] DAC, 2021.	Peking University	--
{Hardware/Software}{Co-Programmable} Framework for Computational {SSDs} to Accelerate Deep Learning Service on {Large-Scale} Graphs[C] FAST, 2022.	KAIST	--
Achieving full parallelism in LSTM via a unified accelerator design[C] ICCD, 2020.	University of Pittsburgh	--
Pasgcn: An reram-based pim design for gcn with adaptively sparsified graphs[J]. TCAD, 2022.	Shanghai Jiao Tong University	--

3.3.2. Hardware Optimization

Title & Basic Information	Affiliation	Code
Ncpu: An embedded neural cpu architecture on resource-constrained low power devices for real-time end-to-end performance[C] MICRO, 2020.	Northwestern Univeristy, Evanston, IL	--
Reduct: Keep it close, keep it cool!: Efficient scaling of dnn inference on multi-core cpus with near-cache compute[C] ISCA, 2021.	ETH Zurich	--
FARNN: FPGA-GPU hybrid acceleration platform for recurrent neural networks[J]. TPDS, 2021.	Sungkyunkwan University	--
Apgan: Approximate gan for robust low energy learning from imprecise components[J]. IEEE Trans. on Computers, 2019.	University of Central Florida	--
An FPGA overlay for CNN inference with fine-grained flexible parallelism[J]. TACO, 2022.	International Institute of Information Technology	--
Pipelined data-parallel CPU/GPU scheduling for multi-DNN real-time inference[C] RTSS, 2019.	University of California, Riverside	--
Deadline-based scheduling for GPU with preemption support[C] RTSS, 2018.	University of Modena and Reggio Emilia Modena	--
Energon: Toward Efficient Acceleration of Transformers Using Dynamic Sparse Attention[J]. TCAD, 2022.	Peking University	--
Light-OPU: An FPGA-based overlay processor for lightweight convolutional neural networks[C] FPGA, 2022.	University of california, Los Angeles	--
Fluid Batching: Exit-Aware Preemptive Serving of Early-Exit Neural Networks on Edge NPUs[J]. arXiv, 2022.	Samsung AI Center, Cambridge	--
BitSystolic: A 26.7 TOPS/W 2b~ 8b NPU with configurable data flows for edge devices[J]. IEEE Trans. on Circuits and Systems I: Regular Papers, 2020.	Duke University	--
PL-NPU: An Energy-Efficient Edge-Device DNN Training Processor With Posit-Based Logarithm-Domain Computing[J]. IEEE Trans. on Circuits and Systems I: Regular Papers, 2022.	Tsinghua University	--