Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware reso...Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.展开更多
Channel pruning can reduce memory consumption and running time with least performance damage,and is one of the most important techniques in network compression.However,existing channel pruning methods mainly focus on ...Channel pruning can reduce memory consumption and running time with least performance damage,and is one of the most important techniques in network compression.However,existing channel pruning methods mainly focus on the pruning of standard convolutional networks,and they rely intensively on time-consuming fine-tuning to achieve the performance improvement.To this end,we present a novel efficient probability-based channel pruning method for depthwise separable convolutional networks.Our method leverages a new simple yet effective probability-based channel pruning criterion by taking the scaling and shifting factors of batch normalization layers into consideration.A novel shifting factor fusion technique is further developed to improve the performance of the pruned networks without requiring extra time-consuming fine-tuning.We apply the proposed method to five representative deep learning networks,namely MobileNetV1,MobileNetV2,ShuffleNetV1,ShuffleNetV2,and GhostNet,to demonstrate the efficiency of our pruning method.Extensive experimental results and comparisons on publicly available CIFAR10,CIFAR100,and ImageNet datasets validate the feasibility of the proposed method.展开更多
Memristor-based neuromorphic computing shows great potential for high-speed and high-throughput signal processing applications,such as electroencephalogram(EEG)signal processing.Nonetheless,the size of one-transistor ...Memristor-based neuromorphic computing shows great potential for high-speed and high-throughput signal processing applications,such as electroencephalogram(EEG)signal processing.Nonetheless,the size of one-transistor one-resistor(1T1R)memristor arrays is limited by the non-ideality of the devices,which prevents the hardware implementation of large and complex networks.In this work,we propose the depthwise separable convolution and bidirectional gate recurrent unit(DSC-BiGRU)network,a lightweight and highly robust hybrid neural network based on 1T1R arrays that enables efficient processing of EEG signals in the temporal,frequency and spatial domains by hybridizing DSC and BiGRU blocks.The network size is reduced and the network robustness is improved while ensuring the network classification accuracy.In the simulation,the measured non-idealities of the 1T1R array are brought into the network through statistical analysis.Compared with traditional convolutional networks,the network parameters are reduced by 95%and the network classification accuracy is improved by 21%at a 95%array yield rate and 5%tolerable error.This work demonstrates that lightweight and highly robust networks based on memristor arrays hold great promise for applications that rely on low consumption and high efficiency.展开更多
文摘Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.
基金the National Natural Science Foundation of China under Grant Nos.62036010 and 62072340the Zhejiang Provincial Natural Science Foundation of China under Grant Nos.LZ21F020001 and LSZ19F020001the Open Project Program of the State Key Laboratory of CAD&CG,Zhejiang University under Grant No.A2220.
文摘Channel pruning can reduce memory consumption and running time with least performance damage,and is one of the most important techniques in network compression.However,existing channel pruning methods mainly focus on the pruning of standard convolutional networks,and they rely intensively on time-consuming fine-tuning to achieve the performance improvement.To this end,we present a novel efficient probability-based channel pruning method for depthwise separable convolutional networks.Our method leverages a new simple yet effective probability-based channel pruning criterion by taking the scaling and shifting factors of batch normalization layers into consideration.A novel shifting factor fusion technique is further developed to improve the performance of the pruned networks without requiring extra time-consuming fine-tuning.We apply the proposed method to five representative deep learning networks,namely MobileNetV1,MobileNetV2,ShuffleNetV1,ShuffleNetV2,and GhostNet,to demonstrate the efficiency of our pruning method.Extensive experimental results and comparisons on publicly available CIFAR10,CIFAR100,and ImageNet datasets validate the feasibility of the proposed method.
基金Project supported by the National Key Research and Development Program of China(Grant No.2019YFB2205102)the National Natural Science Foundation of China(Grant Nos.61974164,62074166,61804181,62004219,62004220,and 62104256).
文摘Memristor-based neuromorphic computing shows great potential for high-speed and high-throughput signal processing applications,such as electroencephalogram(EEG)signal processing.Nonetheless,the size of one-transistor one-resistor(1T1R)memristor arrays is limited by the non-ideality of the devices,which prevents the hardware implementation of large and complex networks.In this work,we propose the depthwise separable convolution and bidirectional gate recurrent unit(DSC-BiGRU)network,a lightweight and highly robust hybrid neural network based on 1T1R arrays that enables efficient processing of EEG signals in the temporal,frequency and spatial domains by hybridizing DSC and BiGRU blocks.The network size is reduced and the network robustness is improved while ensuring the network classification accuracy.In the simulation,the measured non-idealities of the 1T1R array are brought into the network through statistical analysis.Compared with traditional convolutional networks,the network parameters are reduced by 95%and the network classification accuracy is improved by 21%at a 95%array yield rate and 5%tolerable error.This work demonstrates that lightweight and highly robust networks based on memristor arrays hold great promise for applications that rely on low consumption and high efficiency.
