.png){kind=icon}
返回
A novel multi-scale and sparsity auto-encoder for classification
.png){kind=icon}
PRE
AI
DOI：10.1007/s13042-022-01632-5.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
The inspiration for generating the multi-scale feature representation originates from the basic observation that multi-scale is closely related to human visual physiological characteristics. Also, since the increase of hidden layer neurons and the amount of data leads to the rise of redundant information, a large amount of calculation makes a model more complex. This paper proposes a novel learning method, namely, multi-scale feature consistency regularization and L-21-norm minimization sparse auto-encoder (LR21-MSAE). The multi-scale feature consistency regularization can achieve the latent representations and the visual details while retaining multi-scale information. This method ensures that LR21-MSAE can get valid information for better classification accuracy. By implementing the L-21-norm minimization constraint, the LR21-MSAE can adaptively eliminate the potential noise and redundant neurons by enforcing some rows and columns of the weight matrix to be reduced to zero. It can reduce the complexity of the learning model and promote learning sparsity features to generate a compact network. Moreover, introducing the Wasserstein distance in the sparse auto-encoder to measure the difference between the two distributions allows for a more stable training process and faster convergence. To complete the test of the LR21-MSAE model, we choose to conduct the experiments on some publicly available datasets MNIST, Fashion-MNIST, CIFAR-10, USPS, ISOLET, Pendigits, and Ecoli. We demonstrate the advantages of LR21-MSAE, through the experimental results, compared with state-of-the-art feature extraction methods.
Keyword:
Feature representation
Auto-encoder
Multi-scale feature
L-21-norm
Classification
期刊
IF:
2.7
论文数:
3.2K
被引数:
5.6K
机构
引用论文
Multi-information-based convolutional neural network with attention mechanism for pedestrian trajectory prediction基于多信息的注意力卷积神经网络行人轨迹预测
PREAI
分享收藏Context-aware Self-Attention Networks for Natural Language Processing面向自然语言处理的上下文感知自注意网络.png)
NEUROCOMPUTING
NEUROCOMPUTINGIF6.5
PREAI
分享收藏分享收藏Residual wide-kernel deep convolutional auto-encoder for intelligent rotating machinery fault diagnosis with limited samplesNEURAL NETWORKS
NEURAL NETWORKSIF6.3
PREAI
分享收藏A sparse auto-encoder-based deep neural network approach for induction motor faults classificationMEASUREMENT
MEASUREMENTIF5.6
PREAI
分享收藏
查看更多内容