Massive computational complexity and memory requirement of artificial intelligence models impede their deploy-ability on edge computing devices of the Internet of Things(IoT).While Power-of-Two(PoT)quantization is pro...Massive computational complexity and memory requirement of artificial intelligence models impede their deploy-ability on edge computing devices of the Internet of Things(IoT).While Power-of-Two(PoT)quantization is pro-posed to improve the efficiency for edge inference of Deep Neural Networks(DNNs),existing PoT schemes require a huge amount of bit-wise manipulation and have large memory overhead,and their efficiency is bounded by the bottleneck of computation latency and memory footprint.To tackle this challenge,we present an efficient inference approach on the basis of PoT quantization and model compression.An integer-only scalar PoT quantization(IOS-PoT)is designed jointly with a distribution loss regularizer,wherein the regularizer minimizes quantization errors and training disturbances.Additionally,two-stage model compression is developed to effectively reduce memory requirement,and alleviate bandwidth usage in communications of networked heterogenous learning systems.The product look-up table(P-LUT)inference scheme is leveraged to replace bit-shifting with only indexing and addition operations for achieving low-latency computation and implementing efficient edge accelerators.Finally,comprehensive experiments on Residual Networks(ResNets)and efficient architectures with Canadian Institute for Advanced Research(CIFAR),ImageNet,and Real-world Affective Faces Database(RAF-DB)datasets,indicate that our approach achieves 2×∼10×improvement in the reduction of both weight size and computation cost in comparison to state-of-the-art methods.A P-LUT accelerator prototype is implemented on the Xilinx KV260 Field Programmable Gate Array(FPGA)platform for accelerating convolution operations,with performance results showing that P-LUT reduces memory footprint by 1.45×,achieves more than 3×power efficiency and 2×resource efficiency,compared to the conventional bit-shifting scheme.展开更多
The deformation and failure of coal walls in front of a working face cause significant difficulties during mining operations.This study reveals the nonuniform distribution of bearing pressure in front of coal walls ba...The deformation and failure of coal walls in front of a working face cause significant difficulties during mining operations.This study reveals the nonuniform distribution of bearing pressure in front of coal walls based on in situ monitoring data and numerical simulation.Therefore,an eccentric compression mechanical model was established to study the deformation and failure characteristics of a coal wall.The slenderness ratio of the compression bar is introduced to define coal walls.The results showed that instability failure occurs when λ>λ_(c) and material failure occurs when λ≤λ_(c).The instability failure-type coal wall spalling was related to the mining height,eccentricity of roof pressure,the horizontal force,and the reaction moment of the floor.The material failure-type coal wall spalling was related to the cohesion,the internal friction angle of the coal,the upper pressure,and the horizontal force of coal walls.Unstable and destructive coal wall peeling usually occurs at a height of 0.5–0.6 times the mining height,while material damage to coal wall peeling is determined to occur within the range of 0.4-0.6 times the mining depth.The findings contribute to the understanding of the deformation and failure of coal walls.展开更多
With the rapid development of deep learning,neural network models have achieved remarkable performance.However,their large scale and high computational demands still limit widespread deployment.Therefore,model compres...With the rapid development of deep learning,neural network models have achieved remarkable performance.However,their large scale and high computational demands still limit widespread deployment.Therefore,model compression techniques have emerged,aiming to reduce computational complexity,memory usage,and energy overhead while meeting practical deployment needs without sacrificing model performance.This paper provides a systematic overview of recent advances in model compression,with particular emphasis on large-scale models such as vision-language models(VLMs)and large language models(LLMs).We firstly clarify the core concepts and objectives of model compression,then outline persistent challenges,including limited adaptability across different model variants and the trade-off between efficiency and accuracy.We examine five mainstream compression methods in detail:pruning,quantization,knowledge distillation,low-rank factorization,and parameter sharing.For each method,we analyze the guiding principles,strengths,weaknesses,and representative application scenarios.We also present a comparative analysis of trade-offs among compression ratio,accuracy retention,and computational efficiency.Moreover,we review commonly-used benchmarks such as Image Net,Wiki Text,GLUE,and MMLU,along with metrics for evaluating effectiveness in both vision and language tasks.Finally,we outline promising future directions,including automated pipelines,hybrid strategies,hardware-aware optimization,and cross-domain adaptability.This survey provides a comprehensive overview and roadmap for advancing model compression research.展开更多
Activation pruning reduces neural network complexity by eliminating low-importance neuron activations,yet identifying the critical pruning threshold—beyond which accuracy rapidly deteriorates—remains computationally...Activation pruning reduces neural network complexity by eliminating low-importance neuron activations,yet identifying the critical pruning threshold—beyond which accuracy rapidly deteriorates—remains computationally expensive and typically requires exhaustive search.We introduce a thermodynamics-inspired framework that treats activation distributions as energy-filtered physical systems and employs the free energy of activations as a principled evaluation metric.Phase-transition-like phenomena in the free-energy profile—such as extrema,inflection points,and curvature changes—yield reliable estimates of the critical pruning threshold,providing a theoretically grounded means of predicting sharp accuracy degradation.To further enhance efficiency,we propose a renormalized free energy technique that approximates full-evaluation free energy using only the activation distribution of the unpruned network.This eliminates repeated forward passes,dramatically reducing computational overhead and achieving speedups of up to 550×for MLPs.Extensive experiments across diverse vision architectures(MLP,CNN,ResNet,MobileNet,Vision Transformer)and text models(LSTM,BERT,ELECTRA,T5,GPT-2)on multiple datasets validate the generality,robustness,and computational efficiency of our approach.Overall,this work establishes a theoretically grounded and practically effective framework for activation pruning,bridging the gap between analytical understanding and efficient deployment of sparse neural networks.展开更多
Dear Editor,The letter proposes a tensor low-rank orthogonal compression(TLOC)model for a convolutional neural network(CNN),which facilitates its efficient and highly-accurate low-rank representation.Model compression...Dear Editor,The letter proposes a tensor low-rank orthogonal compression(TLOC)model for a convolutional neural network(CNN),which facilitates its efficient and highly-accurate low-rank representation.Model compression is crucial for deploying deep neural network(DNN)models on resource-constrained embedded devices.展开更多
This study investigates the volumetric behaviors of various soils during freeze-thaw(FT)cycles and subsequent one-dimensional(1D)compression from experimental and theoretical studies.Experimental studies were performe...This study investigates the volumetric behaviors of various soils during freeze-thaw(FT)cycles and subsequent one-dimensional(1D)compression from experimental and theoretical studies.Experimental studies were performed on saturated expansive soil specimens with varying compaction conditions and soil structures under different stress states.Experimental results demonstrate that the specimens expand during freezing and contract during thawing.All specimens converge to the same residual void ratio after seven FT cycles,irrespective of their different initial void ratio,stress state,and soil structure.The compression index of the expansive soil specimens increases with the initial void ratio,whereas their swelling index remains nearly constant.A model extending the disturbed state concept(DSC)is proposed to predict the 1D compression behaviors of FT-impacted soils.The model incorporates a parameter,b,to account for the impacts of FT cycles.Empirical equations have been developed to link the key model parameters(i.e.the normalized yield stress and parameter b)to the soil state parameter(i.e.the normalized void ratio)in order to simplify the prediction approach.The proposed model well predicts the results of the tested expansive soil.In addition,the model’s feasibility for other types of soils,including low-and high-plastic clays,and high-plastic organic soils,has been validated using published data from the literature.The proposed model is simple yet reliable for predicting the compression behaviors of soils subjected to FT cycles.展开更多
The steel plate is one of the main products in steel industries,and its surface quality directly affects the final product performance.How to detect surface defects of steel plates in real time during the production p...The steel plate is one of the main products in steel industries,and its surface quality directly affects the final product performance.How to detect surface defects of steel plates in real time during the production process is a challenging problem.The single or fixed model compression method cannot be directly applied to the detection of steel surface defects,because it is difficult to consider the diversity of production tasks,the uncertainty caused by environmental factors,such as communication networks,and the influence of process and working conditions in steel plate production.In this paper,we propose an adaptive model compression method for steel surface defect online detection based on expert knowledge and working conditions.First,we establish an expert system to give lightweight model parameters based on the correlation between defect types and manufacturing processes.Then,lightweight model parameters are adaptively adjusted according to working conditions,which improves detection accuracy while ensuring real-time performance.The experimental results show that compared with the detection method of constant lightweight parameter model,the proposed method makes the total detection time cut down by 23.1%,and the deadline satisfaction ratio increased by 36.5%,while upgrading the accuracy by 4.2%and reducing the false detection rate by 4.3%.展开更多
Automated machine learning(AutoML)pruning methods aim at searching for a pruning strategy automatically to reduce the computational complexity of deep convolutional neural networks(deep CNNs).However,some previous wor...Automated machine learning(AutoML)pruning methods aim at searching for a pruning strategy automatically to reduce the computational complexity of deep convolutional neural networks(deep CNNs).However,some previous work found that the results of many Auto-ML pruning methods cannot even surpass the results of the uniformly pruning method.In this paper,the ineffectiveness of Auto-ML pruning,which is caused by unfull and unfair training of the supernet,is shown.A deep supernet suffers from unfull training because it contains too many candidates.To overcome the unfull training,a stage-wise pruning(SWP)method is proposed,which splits a deep supernet into several stage-wise supernets to reduce the candidate number and utilize inplace distillation to supervise the stage training.Besides,a wide supernet is hit by unfair training since the sampling probability of each channel is unequal.Therefore,the fullnet and the tinynet are sampled in each training iteration to ensure that each channel can be overtrained.Remarkably,the proxy performance of the subnets trained with SWP is closer to the actual performance than that of most of the previous AutoML pruning work.Furthermore,experiments show that SWP achieves the state-of-the-art in both CIFAR-10 and ImageNet under the mobile setting.展开更多
Facial emotion recognition achieves great success with the help of large neural models but also fails to be applied in practical situations due to the large model size of neural methods.To bridge this gap,in this pape...Facial emotion recognition achieves great success with the help of large neural models but also fails to be applied in practical situations due to the large model size of neural methods.To bridge this gap,in this paper,we combine two mainstream model compression methods(pruning and quantization)together,and propose a pruningthen-quantization framework to compress the neural models for facial emotion recognition tasks.Experiments on three datasets show that our model could achieve a high model compression ratio and maintain the model’s high performance well.Besides,We analyze the layer-wise compression performance of our proposed framework to explore its effect and adaptability in fine-grained modules.展开更多
Edge Machine Learning(EdgeML)and Tiny Machine Learning(TinyML)are fast-growing fields that bring machine learning to resource-constrained devices,allowing real-time data processing and decision-making at the network’...Edge Machine Learning(EdgeML)and Tiny Machine Learning(TinyML)are fast-growing fields that bring machine learning to resource-constrained devices,allowing real-time data processing and decision-making at the network’s edge.However,the complexity of model conversion techniques,diverse inference mechanisms,and varied learning strategies make designing and deploying these models challenging.Additionally,deploying TinyML models on resource-constrained hardware with specific software frameworks has broadened EdgeML’s applications across various sectors.These factors underscore the necessity for a comprehensive literature review,as current reviews do not systematically encompass the most recent findings on these topics.Consequently,it provides a comprehensive overview of state-of-the-art techniques in model conversion,inference mechanisms,learning strategies within EdgeML,and deploying these models on resource-constrained edge devices using TinyML.It identifies 90 research articles published between 2018 and 2025,categorizing them into two main areas:(1)model conversion,inference,and learning strategies in EdgeML and(2)deploying TinyML models on resource-constrained hardware using specific software frameworks.In the first category,the synthesis of selected research articles compares and critically reviews various model conversion techniques,inference mechanisms,and learning strategies.In the second category,the synthesis identifies and elaborates on major development boards,software frameworks,sensors,and algorithms used in various applications across six major sectors.As a result,this article provides valuable insights for researchers,practitioners,and developers.It assists them in choosing suitable model conversion techniques,inference mechanisms,learning strategies,hardware development boards,software frameworks,sensors,and algorithms tailored to their specific needs and applications across various sectors.展开更多
BACKGROUND： Varying degrees of inflammatory responses occur during lumbar nerve root compression. Studies have shown that nitric oxide synthase （NOS） and calcitonin gene-related peptide （CGRP） are involved in sec...BACKGROUND： Varying degrees of inflammatory responses occur during lumbar nerve root compression. Studies have shown that nitric oxide synthase （NOS） and calcitonin gene-related peptide （CGRP） are involved in secondary disc inflammation. OBJECTIVE： To observe the effects of warm acupuncture on the ultrastructure of inflammatory mediators in a rat model of lumbar nerve root compression, including NOS and CGRP contents. DESIGN, TIME AND SETTING： Randomized, controlled study, with molecular biological analysis, was performed at the Experimental Center, Sixth People＇s Hospital Affiliated to Shanghai Jiao Tong University, between September 2006 and April 2007. MATERIALS： Acupuncture needles and refined Moxa grains were purchased from Shanghai Taicheng Technology Development Co., Ltd., China; Mobic tablets were purchased from Shanghai Boehringer Ingelheim Pharmaceuticals Co., Ltd., China; enzyme linked immunosorbent assay （ELISA） kits for NOS and CGRP were purchased from ADL Biotechnology, Inc., USA. METHODS： A total of 50, healthy, adult Sprague-Dawley rats, were randomly divided into five groups normal, model, warm acupuncture, acupuncture, and drug, with 10 rats in each group. Rats in the four groups, excluding the normal group, were used to establish models of lumbar nerve root compression. After 3 days, Jiaji points were set using reinforcing-reducing manipulation in the warm acupuncture group. Moxa grains were burned on each needle, with 2 grains each daily. The acupuncture group was the same as the warm acupuncture group, with the exception of non-moxibustion. Mobic suspension （3.75 mg/kg） was used in the oral drug group, once a day. Treatment of each group lasted for 14 consecutive days. Modeling and medication were not performed in the normal group. MAIN OUTCOME MEASURES： The ultrastructure of damaged nerve roots was observed with transmission electron microscopy; NOS and CGRP contents were measured using ELISA. RESULTS： The changes of the radicular ultramicrostructure were characterized by Wallerian degeneration; nerve fibers were clearly demyelinated; axons collapsed or degenerated; outer Schwann cell cytoplasm was swollen and its nucleus was compacted. Compared with the normal group, NOS and CGRP contents in the nerve root compression zone in the model group were significantly increased （P 〈 0.01）. Nerve root edema was improved in the drug, acupuncture and the warm acupuncture groups over the model group. NOS and CGRP expressions were also decreased with the warm acupuncture group having the lowest concentration （P 〈 0.01）. CONCLUSION： In comparison to the known effects of Mobic drug and acupuncture treatments, the warm acupuncture significantly decreased NOS and CGRP expression which helped improve the ultrastructure of the compressed nerve root.展开更多
A novel deep neural network compression model for airport object detection has been presented.This novel model aims at disadvantages of deep neural network,i.e.the complexity of the model and the great cost of calcula...A novel deep neural network compression model for airport object detection has been presented.This novel model aims at disadvantages of deep neural network,i.e.the complexity of the model and the great cost of calculation.According to the requirement of airport object detection,the model obtains temporal and spatial semantic rules from the uncompressed model.These spatial semantic rules are added to the model after parameter compression to assist the detection.The rules can improve the accuracy of the detection model in order to make up for the loss caused by parameter compression.The experiments show that the effect of the novel compression detection model is no worse than that of the uncompressed original model.Even some of the original model false detection can be eliminated through the prior knowledge.展开更多
A numerical simulation of shock wave turbulent boundary layer interaction induced by a 24° compression corner based on Gao-Yong compressible turbulence model was presented.The convection terms and the diffusion t...A numerical simulation of shock wave turbulent boundary layer interaction induced by a 24° compression corner based on Gao-Yong compressible turbulence model was presented.The convection terms and the diffusion terms were calculated using the second-order AUSM(advection upstream splitting method) scheme and the second-order central difference scheme,respectively.The Runge-Kutta time marching method was employed to solve the governing equations for steady state solutions.Significant flow separation-region which indicates highly non-isotropic turbulence structure has been found in the present work due to intensity interaction under the 24° compression corner.Comparisons between the calculated results and experimental data have been carried out,including surface pressure distribution,boundary-layer static pressure profiles and mean velocity profiles.The numerical results agree well with the experimental values,which indicate Gao-Yong compressible turbulence model is suitable for the prediction of shock wave turbulent boundary layer interaction in two-dimensional compression corner flows.展开更多
Modeling of a centrifugal compressor is of great significance to surge characteristics and fluid dynamics in the Altitude Ground Test Facilities(AGTF).Real-time Modular Dynamic System Greitzer(MDSG)modeling for dynami...Modeling of a centrifugal compressor is of great significance to surge characteristics and fluid dynamics in the Altitude Ground Test Facilities(AGTF).Real-time Modular Dynamic System Greitzer(MDSG)modeling for dynamic response and simulation of the compression system is introduced.The centrifugal compressor,pipeline network,and valve are divided into pressure output type and mass flow output type for module modeling,and the two types of components alternate when the system is established.The pressure loss and thermodynamics of the system are considered.An air supply compression system of AGTF is modeled and simulated by the MDSG model.The simulation results of mass flow,pressure,and temperature are compared with the experimental results,and the error is less than 5%,which demonstrates the reliability,practicability,and universality of the MDSG model.展开更多
Warm rotary draw bending provides a feasible method to form the large-diameter thin-walled（LDTW）TC4 bent tubes, which are widely used in the pneumatic system of aircrafts. An accurate prediction of flow behavior of ...Warm rotary draw bending provides a feasible method to form the large-diameter thin-walled（LDTW）TC4 bent tubes, which are widely used in the pneumatic system of aircrafts. An accurate prediction of flow behavior of TC4 tubes considering the couple effects of temperature,strain rate and strain is critical for understanding the deformation behavior of metals and optimizing the processing parameters in warm rotary draw bending of TC4 tubes. In this study, isothermal compression tests of TC4 tube alloy were performed from 573 to 873 K with an interval of 100 K and strain rates of 0.001, 0.010 and0.100 s^（-1）. The prediction of flow behavior was done using two constitutive models, namely modified Arrhenius model and artificial neural network（ANN） model. The predictions of these constitutive models were compared using statistical measures like correlation coefficient（R）, average absolute relative error（AARE） and its variation with the deformation parameters（temperature, strain rate and strain）. Analysis of statistical measures reveals that the two models show high predicted accuracy in terms of R and AARE. Comparatively speaking, the ANN model presents higher predicted accuracy than the modified Arrhenius model. In addition, the predicted accuracy of ANN model presents high stability at the whole deformation parameter ranges, whereas the predictability of the modified Arrhenius model has some fluctuation at different deformation conditions. It presents higher predicted accuracy at temperatures of 573-773 K, strain rates of 0.010-0.100 s^（-1）and strain of 0.04-0.32, while low accuracy at temperature of 873 K, strain rates of 0.001 s^（-1）and strain of 0.36-0.48.Thus, the application of modified Arrhenius model is limited by its relatively low predicted accuracy at some deformation conditions, while the ANN model presents very high predicted accuracy at all deformation conditions,which can be used to study the compression behavior of TC4 tube at the temperature range of 573-873 K and the strain rate of 0.001-0.100 s^（-1）. It can provide guideline for the design of processing parameters in warm rotary draw bending of LDTW TC4 tubes.展开更多
Reconfigurable array architecture has become an important hardware platform for edge-side deployment of convolutional neural networks due to their high parallelism and flexible programmability.However,traditional mult...Reconfigurable array architecture has become an important hardware platform for edge-side deployment of convolutional neural networks due to their high parallelism and flexible programmability.However,traditional multi-branch convolutional networks suffer from computational redundancy,high memory access overhead,and inefficient branch fusion.Therefore,this paper proposes an adaptive multi-branch convolutional module(AMBC)that integrates software-hardware co-optimization.During training,the learnable fusion coefficients are introduced to enable adaptive fusion of multi-scale features,while in the inference phase,the multiple branches and their normalization parameters are merged with the fusion coefficients into a single 3×3 convolutional kernel through operator fusion.On the SIREA-288 reconfigurable platform,compared with unoptimized multi-branch networks,the proposed AMBC reduces external memory accesses by 47.91%and inference latency by 47.20%,achieving a 1.90×speedup.This approach maximizes the utilization of the reconfigurable logic while minimizing both reconfiguration and data-movement overheads in edge inference.展开更多
Accurate prediction of environmental temperature is pivotal for promoting sustainable crop growth.At present,the most effective temperature sensing and prediction system is the Agricultural Internet of Things(AIoT),wh...Accurate prediction of environmental temperature is pivotal for promoting sustainable crop growth.At present,the most effective temperature sensing and prediction system is the Agricultural Internet of Things(AIoT),which deploys a large number of sensors to collect meteorological data and transmits them to the cloud server for prediction.However,this procedure is computationally and communicationally expensive for resourceconstrained AIoT.Recently,Semantic Communication(SC)has shown potential in efficient data transmission,but existing methods overlook the repetitive semantic information whilst sensing data,bringing additional overheads.With the resource-constraint nature of AIoT in mind,we propose the Semantic Communication-enabled Cognitive Agriculture Framework(SC-CAF)for delivering accurate temperature predictions.The proposed SC-CAF incorporates an intelligent analysis layer that performs the temperature prediction and model training and distribution,while a semantic layer transmitting the semantic information extracted from raw data based on the download model,ultimately to reduce communication overheads in AIoT.Furthermore,we propose a novel model called the Light Temperature Semantic Communication(LTSC)by adopting skip-attention and semantic compressor to avoid unnecessary computation and repetitive information,thereby addressing the semantic redundancy issues in sensing data.We also develop a Semantic-based Model Compression(SCMC)algorithm to alleviate the computation and bandwidth burden,enabling AIoT to explore the extensive usage of SC.Experimental results demonstrate that the proposed SC-CAF achieves the lowest prediction error while reducing Floating Point Operations(FLOPs)by 95.88%,memory requirements by 78.30%,Graphics Processing Unit(GPU)power by 50.77%,and time latency by 84.44%,outperforming notable state-of-the-art methods.展开更多
The high temperature split Hopkinson pressure bar （SHPB） compression experiment is conducted to obtain the data relationship among strain, strain rate and flow stress from room temperature to 550 C for aeronautical ...The high temperature split Hopkinson pressure bar （SHPB） compression experiment is conducted to obtain the data relationship among strain, strain rate and flow stress from room temperature to 550 C for aeronautical aluminum alloy 7050-T7451. Combined high-speed orthogonal cutting experiments with the cutting process simulations, the data relationship of high temperature, high strain rate and large strain in high-speed cutting is modified. The Johnson-Cook empirical model considering the effects of strain hardening, strain rate hardening and thermal softening is selected to describe the data relationship in high-speed cutting, and the material constants of flow stress constitutive model for aluminum alloy 7050-T7451 are determined. Finally, the constitutive model of aluminum alloy 7050-T7451 is established through experiment and simulation verification in high-speed cutting. The model is proved to be reasonable by matching the measured values of the cutting force with the estimated results from FEM simulations.展开更多
Recent advancements in natural language processing have given rise to numerous pre-training language models in question-answering systems.However,with the constant evolution of algorithms,data,and computing power,the ...Recent advancements in natural language processing have given rise to numerous pre-training language models in question-answering systems.However,with the constant evolution of algorithms,data,and computing power,the increasing size and complexity of these models have led to increased training costs and reduced efficiency.This study aims to minimize the inference time of such models while maintaining computational performance.It also proposes a novel Distillation model for PAL-BERT(DPAL-BERT),specifically,employs knowledge distillation,using the PAL-BERT model as the teacher model to train two student models:DPAL-BERT-Bi and DPAL-BERTC.This research enhances the dataset through techniques such as masking,replacement,and n-gram sampling to optimize knowledge transfer.The experimental results showed that the distilled models greatly outperform models trained from scratch.In addition,although the distilled models exhibit a slight decrease in performance compared to PAL-BERT,they significantly reduce inference time to just 0.25%of the original.This demonstrates the effectiveness of the proposed approach in balancing model performance and efficiency.展开更多
Deep learning technology has been widely used in computer vision,speech recognition,natural language processing,and other related fields.The deep learning algorithm has high precision and high reliability.However,the ...Deep learning technology has been widely used in computer vision,speech recognition,natural language processing,and other related fields.The deep learning algorithm has high precision and high reliability.However,the lack of resources in the edge terminal equipment makes it difficult to run deep learning algorithms that require more memory and computing power.In this paper,we propose MoTransFrame,a general model processing framework for deep learning models.Instead of designing a model compression algorithm with a high compression ratio,MoTransFrame can transplant popular convolutional neural networks models to resources-starved edge devices promptly and accurately.By the integration method,Deep learning models can be converted into portable projects for Arduino,a typical edge device with limited resources.Our experiments show that MoTransFrame has good adaptability in edge devices with limited memories.It is more flexible than other model transplantation methods.It can keep a small loss of model accuracy when the number of parameters is compressed by tens of times.At the same time,the computational resources needed in the reasoning process are less than what the edge node could handle.展开更多
基金This work was supported by Open Fund Project of State Key Laboratory of Intelligent Vehicle Safety Technology by Grant with No.IVSTSKL-202311Key Projects of Science and Technology Research Programme of Chongqing Municipal Education Commission by Grant with No.KJZD-K202301505+1 种基金Cooperation Project between Chongqing Municipal Undergraduate Universities and Institutes Affiliated to the Chinese Academy of Sciences in 2021 by Grant with No.HZ2021015Chongqing Graduate Student Research Innovation Program by Grant with No.CYS240801.
文摘Massive computational complexity and memory requirement of artificial intelligence models impede their deploy-ability on edge computing devices of the Internet of Things(IoT).While Power-of-Two(PoT)quantization is pro-posed to improve the efficiency for edge inference of Deep Neural Networks(DNNs),existing PoT schemes require a huge amount of bit-wise manipulation and have large memory overhead,and their efficiency is bounded by the bottleneck of computation latency and memory footprint.To tackle this challenge,we present an efficient inference approach on the basis of PoT quantization and model compression.An integer-only scalar PoT quantization(IOS-PoT)is designed jointly with a distribution loss regularizer,wherein the regularizer minimizes quantization errors and training disturbances.Additionally,two-stage model compression is developed to effectively reduce memory requirement,and alleviate bandwidth usage in communications of networked heterogenous learning systems.The product look-up table(P-LUT)inference scheme is leveraged to replace bit-shifting with only indexing and addition operations for achieving low-latency computation and implementing efficient edge accelerators.Finally,comprehensive experiments on Residual Networks(ResNets)and efficient architectures with Canadian Institute for Advanced Research(CIFAR),ImageNet,and Real-world Affective Faces Database(RAF-DB)datasets,indicate that our approach achieves 2×∼10×improvement in the reduction of both weight size and computation cost in comparison to state-of-the-art methods.A P-LUT accelerator prototype is implemented on the Xilinx KV260 Field Programmable Gate Array(FPGA)platform for accelerating convolution operations,with performance results showing that P-LUT reduces memory footprint by 1.45×,achieves more than 3×power efficiency and 2×resource efficiency,compared to the conventional bit-shifting scheme.
基金Youth Innovation Team of Shandong Higher Education Institutions,Grant/Award Number:2022KJ214Shandong Postdoctoral Science Foundation,Grant/Award Number:SDCXZG‐202303031+2 种基金China Postdoctoral Science Foundation,Grant/Award Number:2023M732109National Natural Science Foundation of China,Grant/Award Number:52209141Natural Science Foundation of Shandong Province,China,Grant/Award Number:ZR2021QE069。
文摘The deformation and failure of coal walls in front of a working face cause significant difficulties during mining operations.This study reveals the nonuniform distribution of bearing pressure in front of coal walls based on in situ monitoring data and numerical simulation.Therefore,an eccentric compression mechanical model was established to study the deformation and failure characteristics of a coal wall.The slenderness ratio of the compression bar is introduced to define coal walls.The results showed that instability failure occurs when λ>λ_(c) and material failure occurs when λ≤λ_(c).The instability failure-type coal wall spalling was related to the mining height,eccentricity of roof pressure,the horizontal force,and the reaction moment of the floor.The material failure-type coal wall spalling was related to the cohesion,the internal friction angle of the coal,the upper pressure,and the horizontal force of coal walls.Unstable and destructive coal wall peeling usually occurs at a height of 0.5–0.6 times the mining height,while material damage to coal wall peeling is determined to occur within the range of 0.4-0.6 times the mining depth.The findings contribute to the understanding of the deformation and failure of coal walls.
文摘With the rapid development of deep learning,neural network models have achieved remarkable performance.However,their large scale and high computational demands still limit widespread deployment.Therefore,model compression techniques have emerged,aiming to reduce computational complexity,memory usage,and energy overhead while meeting practical deployment needs without sacrificing model performance.This paper provides a systematic overview of recent advances in model compression,with particular emphasis on large-scale models such as vision-language models(VLMs)and large language models(LLMs).We firstly clarify the core concepts and objectives of model compression,then outline persistent challenges,including limited adaptability across different model variants and the trade-off between efficiency and accuracy.We examine five mainstream compression methods in detail:pruning,quantization,knowledge distillation,low-rank factorization,and parameter sharing.For each method,we analyze the guiding principles,strengths,weaknesses,and representative application scenarios.We also present a comparative analysis of trade-offs among compression ratio,accuracy retention,and computational efficiency.Moreover,we review commonly-used benchmarks such as Image Net,Wiki Text,GLUE,and MMLU,along with metrics for evaluating effectiveness in both vision and language tasks.Finally,we outline promising future directions,including automated pipelines,hybrid strategies,hardware-aware optimization,and cross-domain adaptability.This survey provides a comprehensive overview and roadmap for advancing model compression research.
基金output of a research project implemented as part of the Basic Research Program at HSE University。
文摘Activation pruning reduces neural network complexity by eliminating low-importance neuron activations,yet identifying the critical pruning threshold—beyond which accuracy rapidly deteriorates—remains computationally expensive and typically requires exhaustive search.We introduce a thermodynamics-inspired framework that treats activation distributions as energy-filtered physical systems and employs the free energy of activations as a principled evaluation metric.Phase-transition-like phenomena in the free-energy profile—such as extrema,inflection points,and curvature changes—yield reliable estimates of the critical pruning threshold,providing a theoretically grounded means of predicting sharp accuracy degradation.To further enhance efficiency,we propose a renormalized free energy technique that approximates full-evaluation free energy using only the activation distribution of the unpruned network.This eliminates repeated forward passes,dramatically reducing computational overhead and achieving speedups of up to 550×for MLPs.Extensive experiments across diverse vision architectures(MLP,CNN,ResNet,MobileNet,Vision Transformer)and text models(LSTM,BERT,ELECTRA,T5,GPT-2)on multiple datasets validate the generality,robustness,and computational efficiency of our approach.Overall,this work establishes a theoretically grounded and practically effective framework for activation pruning,bridging the gap between analytical understanding and efficient deployment of sparse neural networks.
基金supported by the Science and Technology Innovation Key R&D Program of Chongqing(CSTB2025TIAD-STX0032)National Key Research and Development Program of China(2024YFF0908200)+1 种基金the Chongqing Technology Innovation and Application Development Special Key Project(CSTB2024TIAD-KPX0018)the Southwest University Graduate Student Research Innovation(SWUB24051)。
文摘Dear Editor,The letter proposes a tensor low-rank orthogonal compression(TLOC)model for a convolutional neural network(CNN),which facilitates its efficient and highly-accurate low-rank representation.Model compression is crucial for deploying deep neural network(DNN)models on resource-constrained embedded devices.
基金support from the Natural Sciences and Engineering Research Council of Canada(NSERC)through the Discovery Grant(Grant No.5808)received in 2019 for his research programsThe third author appreciates the funding from the National Natural Science Foundation of China(Grant No.52378365)Hubei Key Research&Development Program(Grant No.2023BCB112).
文摘This study investigates the volumetric behaviors of various soils during freeze-thaw(FT)cycles and subsequent one-dimensional(1D)compression from experimental and theoretical studies.Experimental studies were performed on saturated expansive soil specimens with varying compaction conditions and soil structures under different stress states.Experimental results demonstrate that the specimens expand during freezing and contract during thawing.All specimens converge to the same residual void ratio after seven FT cycles,irrespective of their different initial void ratio,stress state,and soil structure.The compression index of the expansive soil specimens increases with the initial void ratio,whereas their swelling index remains nearly constant.A model extending the disturbed state concept(DSC)is proposed to predict the 1D compression behaviors of FT-impacted soils.The model incorporates a parameter,b,to account for the impacts of FT cycles.Empirical equations have been developed to link the key model parameters(i.e.the normalized yield stress and parameter b)to the soil state parameter(i.e.the normalized void ratio)in order to simplify the prediction approach.The proposed model well predicts the results of the tested expansive soil.In addition,the model’s feasibility for other types of soils,including low-and high-plastic clays,and high-plastic organic soils,has been validated using published data from the literature.The proposed model is simple yet reliable for predicting the compression behaviors of soils subjected to FT cycles.
基金supported by the National Key R&D Program of China(No.2018AAA0100500)the National Natural Science Foundation of China(Nos.62232004 and 61632008)+3 种基金the Jiangsu Provincial Key Laboratory of Network and Information Security(No.BM2003201)the Key Laboratory of Computer Network and Information Integration of Ministry of Education of China(No.93K-9)the Collaborative Innovation Center of Novel Software Technology and Industrializationthe Big Data Computing Center of Southeast University in China for providing the experiment environment and computing facility.
文摘The steel plate is one of the main products in steel industries,and its surface quality directly affects the final product performance.How to detect surface defects of steel plates in real time during the production process is a challenging problem.The single or fixed model compression method cannot be directly applied to the detection of steel surface defects,because it is difficult to consider the diversity of production tasks,the uncertainty caused by environmental factors,such as communication networks,and the influence of process and working conditions in steel plate production.In this paper,we propose an adaptive model compression method for steel surface defect online detection based on expert knowledge and working conditions.First,we establish an expert system to give lightweight model parameters based on the correlation between defect types and manufacturing processes.Then,lightweight model parameters are adaptively adjusted according to working conditions,which improves detection accuracy while ensuring real-time performance.The experimental results show that compared with the detection method of constant lightweight parameter model,the proposed method makes the total detection time cut down by 23.1%,and the deadline satisfaction ratio increased by 36.5%,while upgrading the accuracy by 4.2%and reducing the false detection rate by 4.3%.
基金This work was supported by Natural Science Foundation of Zhejiang Province,China(No.LY21F030018)National Key R&D Program of China(No.2018YFB 1308400).
文摘Automated machine learning(AutoML)pruning methods aim at searching for a pruning strategy automatically to reduce the computational complexity of deep convolutional neural networks(deep CNNs).However,some previous work found that the results of many Auto-ML pruning methods cannot even surpass the results of the uniformly pruning method.In this paper,the ineffectiveness of Auto-ML pruning,which is caused by unfull and unfair training of the supernet,is shown.A deep supernet suffers from unfull training because it contains too many candidates.To overcome the unfull training,a stage-wise pruning(SWP)method is proposed,which splits a deep supernet into several stage-wise supernets to reduce the candidate number and utilize inplace distillation to supervise the stage training.Besides,a wide supernet is hit by unfair training since the sampling probability of each channel is unequal.Therefore,the fullnet and the tinynet are sampled in each training iteration to ensure that each channel can be overtrained.Remarkably,the proxy performance of the subnets trained with SWP is closer to the actual performance than that of most of the previous AutoML pruning work.Furthermore,experiments show that SWP achieves the state-of-the-art in both CIFAR-10 and ImageNet under the mobile setting.
基金supported in part by the Technological Breakthrough Project of Science,Technology and Innovation Commission of Shenzhen Municipality(No.JSGG20201102162000001)InnoHK Initiative of Hong Kong SAR Government,and the Laboratory for AI-Powered Financial Technologies Ltd.
文摘Facial emotion recognition achieves great success with the help of large neural models but also fails to be applied in practical situations due to the large model size of neural methods.To bridge this gap,in this paper,we combine two mainstream model compression methods(pruning and quantization)together,and propose a pruningthen-quantization framework to compress the neural models for facial emotion recognition tasks.Experiments on three datasets show that our model could achieve a high model compression ratio and maintain the model’s high performance well.Besides,We analyze the layer-wise compression performance of our proposed framework to explore its effect and adaptability in fine-grained modules.
文摘Edge Machine Learning(EdgeML)and Tiny Machine Learning(TinyML)are fast-growing fields that bring machine learning to resource-constrained devices,allowing real-time data processing and decision-making at the network’s edge.However,the complexity of model conversion techniques,diverse inference mechanisms,and varied learning strategies make designing and deploying these models challenging.Additionally,deploying TinyML models on resource-constrained hardware with specific software frameworks has broadened EdgeML’s applications across various sectors.These factors underscore the necessity for a comprehensive literature review,as current reviews do not systematically encompass the most recent findings on these topics.Consequently,it provides a comprehensive overview of state-of-the-art techniques in model conversion,inference mechanisms,learning strategies within EdgeML,and deploying these models on resource-constrained edge devices using TinyML.It identifies 90 research articles published between 2018 and 2025,categorizing them into two main areas:(1)model conversion,inference,and learning strategies in EdgeML and(2)deploying TinyML models on resource-constrained hardware using specific software frameworks.In the first category,the synthesis of selected research articles compares and critically reviews various model conversion techniques,inference mechanisms,and learning strategies.In the second category,the synthesis identifies and elaborates on major development boards,software frameworks,sensors,and algorithms used in various applications across six major sectors.As a result,this article provides valuable insights for researchers,practitioners,and developers.It assists them in choosing suitable model conversion techniques,inference mechanisms,learning strategies,hardware development boards,software frameworks,sensors,and algorithms tailored to their specific needs and applications across various sectors.
基金Modern Projects of Traditional Chinese Medicine of Shanghai Science and Technology Commission, No.08DZ1973200Research Projects of Shanghai Bureau of Public Health,No.2006Q004L
文摘BACKGROUND： Varying degrees of inflammatory responses occur during lumbar nerve root compression. Studies have shown that nitric oxide synthase （NOS） and calcitonin gene-related peptide （CGRP） are involved in secondary disc inflammation. OBJECTIVE： To observe the effects of warm acupuncture on the ultrastructure of inflammatory mediators in a rat model of lumbar nerve root compression, including NOS and CGRP contents. DESIGN, TIME AND SETTING： Randomized, controlled study, with molecular biological analysis, was performed at the Experimental Center, Sixth People＇s Hospital Affiliated to Shanghai Jiao Tong University, between September 2006 and April 2007. MATERIALS： Acupuncture needles and refined Moxa grains were purchased from Shanghai Taicheng Technology Development Co., Ltd., China; Mobic tablets were purchased from Shanghai Boehringer Ingelheim Pharmaceuticals Co., Ltd., China; enzyme linked immunosorbent assay （ELISA） kits for NOS and CGRP were purchased from ADL Biotechnology, Inc., USA. METHODS： A total of 50, healthy, adult Sprague-Dawley rats, were randomly divided into five groups normal, model, warm acupuncture, acupuncture, and drug, with 10 rats in each group. Rats in the four groups, excluding the normal group, were used to establish models of lumbar nerve root compression. After 3 days, Jiaji points were set using reinforcing-reducing manipulation in the warm acupuncture group. Moxa grains were burned on each needle, with 2 grains each daily. The acupuncture group was the same as the warm acupuncture group, with the exception of non-moxibustion. Mobic suspension （3.75 mg/kg） was used in the oral drug group, once a day. Treatment of each group lasted for 14 consecutive days. Modeling and medication were not performed in the normal group. MAIN OUTCOME MEASURES： The ultrastructure of damaged nerve roots was observed with transmission electron microscopy; NOS and CGRP contents were measured using ELISA. RESULTS： The changes of the radicular ultramicrostructure were characterized by Wallerian degeneration; nerve fibers were clearly demyelinated; axons collapsed or degenerated; outer Schwann cell cytoplasm was swollen and its nucleus was compacted. Compared with the normal group, NOS and CGRP contents in the nerve root compression zone in the model group were significantly increased （P 〈 0.01）. Nerve root edema was improved in the drug, acupuncture and the warm acupuncture groups over the model group. NOS and CGRP expressions were also decreased with the warm acupuncture group having the lowest concentration （P 〈 0.01）. CONCLUSION： In comparison to the known effects of Mobic drug and acupuncture treatments, the warm acupuncture significantly decreased NOS and CGRP expression which helped improve the ultrastructure of the compressed nerve root.
文摘A novel deep neural network compression model for airport object detection has been presented.This novel model aims at disadvantages of deep neural network,i.e.the complexity of the model and the great cost of calculation.According to the requirement of airport object detection,the model obtains temporal and spatial semantic rules from the uncompressed model.These spatial semantic rules are added to the model after parameter compression to assist the detection.The rules can improve the accuracy of the detection model in order to make up for the loss caused by parameter compression.The experiments show that the effect of the novel compression detection model is no worse than that of the uncompressed original model.Even some of the original model false detection can be eliminated through the prior knowledge.
文摘A numerical simulation of shock wave turbulent boundary layer interaction induced by a 24° compression corner based on Gao-Yong compressible turbulence model was presented.The convection terms and the diffusion terms were calculated using the second-order AUSM(advection upstream splitting method) scheme and the second-order central difference scheme,respectively.The Runge-Kutta time marching method was employed to solve the governing equations for steady state solutions.Significant flow separation-region which indicates highly non-isotropic turbulence structure has been found in the present work due to intensity interaction under the 24° compression corner.Comparisons between the calculated results and experimental data have been carried out,including surface pressure distribution,boundary-layer static pressure profiles and mean velocity profiles.The numerical results agree well with the experimental values,which indicate Gao-Yong compressible turbulence model is suitable for the prediction of shock wave turbulent boundary layer interaction in two-dimensional compression corner flows.
基金supported in part by the Stable Support Research Project of AECC Sichuan Gas Turbine Establishment,China(No.GJCZ-0013-19)the Open Foundation of State Key Laboratory of Compressor Technology,China(Compressor Technology Laboratory of Anhui Province)(No.SKL-YSJ2020007).
文摘Modeling of a centrifugal compressor is of great significance to surge characteristics and fluid dynamics in the Altitude Ground Test Facilities(AGTF).Real-time Modular Dynamic System Greitzer(MDSG)modeling for dynamic response and simulation of the compression system is introduced.The centrifugal compressor,pipeline network,and valve are divided into pressure output type and mass flow output type for module modeling,and the two types of components alternate when the system is established.The pressure loss and thermodynamics of the system are considered.An air supply compression system of AGTF is modeled and simulated by the MDSG model.The simulation results of mass flow,pressure,and temperature are compared with the experimental results,and the error is less than 5%,which demonstrates the reliability,practicability,and universality of the MDSG model.
基金financially supported by the National Natural Science Foundation of China(Nos.51275415 and50905144)the Natural Science Basic Research Plan in Shanxi Province(No.2011JQ6004)the Program of the Ministry of Education of China for Introducing Talents of Discipline to Universities(No.B08040)
文摘Warm rotary draw bending provides a feasible method to form the large-diameter thin-walled（LDTW）TC4 bent tubes, which are widely used in the pneumatic system of aircrafts. An accurate prediction of flow behavior of TC4 tubes considering the couple effects of temperature,strain rate and strain is critical for understanding the deformation behavior of metals and optimizing the processing parameters in warm rotary draw bending of TC4 tubes. In this study, isothermal compression tests of TC4 tube alloy were performed from 573 to 873 K with an interval of 100 K and strain rates of 0.001, 0.010 and0.100 s^（-1）. The prediction of flow behavior was done using two constitutive models, namely modified Arrhenius model and artificial neural network（ANN） model. The predictions of these constitutive models were compared using statistical measures like correlation coefficient（R）, average absolute relative error（AARE） and its variation with the deformation parameters（temperature, strain rate and strain）. Analysis of statistical measures reveals that the two models show high predicted accuracy in terms of R and AARE. Comparatively speaking, the ANN model presents higher predicted accuracy than the modified Arrhenius model. In addition, the predicted accuracy of ANN model presents high stability at the whole deformation parameter ranges, whereas the predictability of the modified Arrhenius model has some fluctuation at different deformation conditions. It presents higher predicted accuracy at temperatures of 573-773 K, strain rates of 0.010-0.100 s^（-1）and strain of 0.04-0.32, while low accuracy at temperature of 873 K, strain rates of 0.001 s^（-1）and strain of 0.36-0.48.Thus, the application of modified Arrhenius model is limited by its relatively low predicted accuracy at some deformation conditions, while the ANN model presents very high predicted accuracy at all deformation conditions,which can be used to study the compression behavior of TC4 tube at the temperature range of 573-873 K and the strain rate of 0.001-0.100 s^（-1）. It can provide guideline for the design of processing parameters in warm rotary draw bending of LDTW TC4 tubes.
基金Supported by the National Science and Technology Major Project of China(2022ZD0119005)the Natural Science Project of Shaanxi Province(2025JC-YBMS-754,2024JC-YBMS-539)。
文摘Reconfigurable array architecture has become an important hardware platform for edge-side deployment of convolutional neural networks due to their high parallelism and flexible programmability.However,traditional multi-branch convolutional networks suffer from computational redundancy,high memory access overhead,and inefficient branch fusion.Therefore,this paper proposes an adaptive multi-branch convolutional module(AMBC)that integrates software-hardware co-optimization.During training,the learnable fusion coefficients are introduced to enable adaptive fusion of multi-scale features,while in the inference phase,the multiple branches and their normalization parameters are merged with the fusion coefficients into a single 3×3 convolutional kernel through operator fusion.On the SIREA-288 reconfigurable platform,compared with unoptimized multi-branch networks,the proposed AMBC reduces external memory accesses by 47.91%and inference latency by 47.20%,achieving a 1.90×speedup.This approach maximizes the utilization of the reconfigurable logic while minimizing both reconfiguration and data-movement overheads in edge inference.
基金supported by the Key Research and Development Project of Hubei Province(No.2024BAB070),China。
文摘Accurate prediction of environmental temperature is pivotal for promoting sustainable crop growth.At present,the most effective temperature sensing and prediction system is the Agricultural Internet of Things(AIoT),which deploys a large number of sensors to collect meteorological data and transmits them to the cloud server for prediction.However,this procedure is computationally and communicationally expensive for resourceconstrained AIoT.Recently,Semantic Communication(SC)has shown potential in efficient data transmission,but existing methods overlook the repetitive semantic information whilst sensing data,bringing additional overheads.With the resource-constraint nature of AIoT in mind,we propose the Semantic Communication-enabled Cognitive Agriculture Framework(SC-CAF)for delivering accurate temperature predictions.The proposed SC-CAF incorporates an intelligent analysis layer that performs the temperature prediction and model training and distribution,while a semantic layer transmitting the semantic information extracted from raw data based on the download model,ultimately to reduce communication overheads in AIoT.Furthermore,we propose a novel model called the Light Temperature Semantic Communication(LTSC)by adopting skip-attention and semantic compressor to avoid unnecessary computation and repetitive information,thereby addressing the semantic redundancy issues in sensing data.We also develop a Semantic-based Model Compression(SCMC)algorithm to alleviate the computation and bandwidth burden,enabling AIoT to explore the extensive usage of SC.Experimental results demonstrate that the proposed SC-CAF achieves the lowest prediction error while reducing Floating Point Operations(FLOPs)by 95.88%,memory requirements by 78.30%,Graphics Processing Unit(GPU)power by 50.77%,and time latency by 84.44%,outperforming notable state-of-the-art methods.
文摘The high temperature split Hopkinson pressure bar （SHPB） compression experiment is conducted to obtain the data relationship among strain, strain rate and flow stress from room temperature to 550 C for aeronautical aluminum alloy 7050-T7451. Combined high-speed orthogonal cutting experiments with the cutting process simulations, the data relationship of high temperature, high strain rate and large strain in high-speed cutting is modified. The Johnson-Cook empirical model considering the effects of strain hardening, strain rate hardening and thermal softening is selected to describe the data relationship in high-speed cutting, and the material constants of flow stress constitutive model for aluminum alloy 7050-T7451 are determined. Finally, the constitutive model of aluminum alloy 7050-T7451 is established through experiment and simulation verification in high-speed cutting. The model is proved to be reasonable by matching the measured values of the cutting force with the estimated results from FEM simulations.
基金supported by Sichuan Science and Technology Program(2023YFSY0026,2023YFH0004).
文摘Recent advancements in natural language processing have given rise to numerous pre-training language models in question-answering systems.However,with the constant evolution of algorithms,data,and computing power,the increasing size and complexity of these models have led to increased training costs and reduced efficiency.This study aims to minimize the inference time of such models while maintaining computational performance.It also proposes a novel Distillation model for PAL-BERT(DPAL-BERT),specifically,employs knowledge distillation,using the PAL-BERT model as the teacher model to train two student models:DPAL-BERT-Bi and DPAL-BERTC.This research enhances the dataset through techniques such as masking,replacement,and n-gram sampling to optimize knowledge transfer.The experimental results showed that the distilled models greatly outperform models trained from scratch.In addition,although the distilled models exhibit a slight decrease in performance compared to PAL-BERT,they significantly reduce inference time to just 0.25%of the original.This demonstrates the effectiveness of the proposed approach in balancing model performance and efficiency.
基金supported by The National Key Research and Development Program of China(2018YFB1800202,2016YFB1000302,SQ2019ZD090149,2018YFB0204301)the CETC Joint Advanced Research Foundation(6141B08080101)+1 种基金The Major Special Science and Technology Project of Hainan Province(ZDKJ2019008)The New Generation of Artificial Intelligence Special Action Project(AI20191125008).
文摘Deep learning technology has been widely used in computer vision,speech recognition,natural language processing,and other related fields.The deep learning algorithm has high precision and high reliability.However,the lack of resources in the edge terminal equipment makes it difficult to run deep learning algorithms that require more memory and computing power.In this paper,we propose MoTransFrame,a general model processing framework for deep learning models.Instead of designing a model compression algorithm with a high compression ratio,MoTransFrame can transplant popular convolutional neural networks models to resources-starved edge devices promptly and accurately.By the integration method,Deep learning models can be converted into portable projects for Arduino,a typical edge device with limited resources.Our experiments show that MoTransFrame has good adaptability in edge devices with limited memories.It is more flexible than other model transplantation methods.It can keep a small loss of model accuracy when the number of parameters is compressed by tens of times.At the same time,the computational resources needed in the reasoning process are less than what the edge node could handle.
