Bias Variance

مقاله A tiny deep capsule network
شبکه ای که به تازگی ارایه شد و قابلیت استفاده در زمینه های متفاوت را دارد. خواندن این مقاله بسیار توصیه می شود.

The capsule network (CapsNet) is a novel network model that can learn spatial information in images. However, the performance of CapsNet on complex datasets (such as CIFAR10) is limited and it requires a large number of parameters. These disadvantages make CapsNet less useful, especially in some resource-constrained devices. To solve this problem, we propose a novel tiny deep capsule architecture (CapsInfor), which consists of many fast tensor capsule layers (FastCaps) with a novel routing process. CapsInfor requires only a few parameters to achieve satisfactory performance. For example, on CIFAR10, the accuracy of CapsInfor is 9.32% higher than that of CapsNet, but the parameters are reduced by 97.53%. CapsInfor is composed of multiple pipelines each of which processes a kind of image information. To achieve information interaction between pipelines, a novel cross node is proposed to implement pipeline-level capsule routing. A new decision maker is used to analyze the predicted values of pipelines and gives the final classification result. Using these proposed methods, CapsInfor achieves competitive results on CIFAR10, CIFAR100, FMNIST, and SVHN. Besides, it is proved that CapsInfor has satisfactory affine robustness on affNIST. To alleviate the problem that the parameter explosion with increasing the number of classes, a novel two-level classification method is proposed. This method can effectively reduce the parameters of the model on the 10 categories and 100 categories tasks. The experimental results confirm that CapsInfor is a tiny deep capsule model with satisfactory classification accuracy and affine robustness.



#یادگیری_عمیق #معرفی_مقاله #capsule_network

🌴 سایت | 🌺 کانال | 🌳 پشتیبانی

SpringerLink

A tiny deep capsule network

International Journal of Machine Learning and Cybernetics - The capsule network (CapsNet) is a novel network model that can learn spatial information in images. However, the performance of CapsNet...

www.tgoop.com/biasvariance_ir/105

308 viewsBias Variance, edited  Oct 4, 2021 at 16:25

مقاله A tiny deep capsule network
شبکه ای که به تازگی ارایه شد و قابلیت استفاده در زمینه های متفاوت را دارد. خواندن این مقاله بسیار توصیه می شود.

The capsule network (CapsNet) is a novel network model that can learn spatial information in images. However, the performance of CapsNet on complex datasets (such as CIFAR10) is limited and it requires a large number of parameters. These disadvantages make CapsNet less useful, especially in some resource-constrained devices. To solve this problem, we propose a novel tiny deep capsule architecture (CapsInfor), which consists of many fast tensor capsule layers (FastCaps) with a novel routing process. CapsInfor requires only a few parameters to achieve satisfactory performance. For example, on CIFAR10, the accuracy of CapsInfor is 9.32% higher than that of CapsNet, but the parameters are reduced by 97.53%. CapsInfor is composed of multiple pipelines each of which processes a kind of image information. To achieve information interaction between pipelines, a novel cross node is proposed to implement pipeline-level capsule routing. A new decision maker is used to analyze the predicted values of pipelines and gives the final classification result. Using these proposed methods, CapsInfor achieves competitive results on CIFAR10, CIFAR100, FMNIST, and SVHN. Besides, it is proved that CapsInfor has satisfactory affine robustness on affNIST. To alleviate the problem that the parameter explosion with increasing the number of classes, a novel two-level classification method is proposed. This method can effectively reduce the parameters of the model on the 10 categories and 100 categories tasks. The experimental results confirm that CapsInfor is a tiny deep capsule model with satisfactory classification accuracy and affine robustness.



#یادگیری_عمیق #معرفی_مقاله #capsule_network

🌴 سایت | 🌺 کانال | 🌳 پشتیبانی

BY Bias Variance