当前位置:网站首页>【Mixed Pooling】《Mixed Pooling for Convolutional Neural Networks》
【Mixed Pooling】《Mixed Pooling for Convolutional Neural Networks》
2022-07-02 06:26:00 【bryant_meng】
RSKT-2014
International conference on rough sets and knowledge technology
文章目录
1 Background and Motivation
池化层的作用(一文看尽深度学习中的9种池化方法!)
- 增大网络感受野
- 抑制噪声,降低信息冗余
- 降低模型计算量,降低网络优化难度,防止网络过拟合
- 使模型对输入图像中的特征位置变化更加鲁棒
作者针对 max 和 ave pooling 的缺点,
提出了 mix pooling——randomly employs the local max pooling and average pooling methods when training CNNs
2 Review of Convolutional Neural Networks
- Convolutional Layer,包括卷积操作和 activation function
- Non-linear Transformation Layer,也即 normalization 层,现在比较流行的是 BN 等,以前的是 LCN(local contrast normalization) 和 AlexNet 的 LRN(the local response normalization) 等,PS:论文中 LCN 的公式感觉有问题,LRN 细节也原版论文也有差距,形式基本一致
- Feature Pooling Layer
3 Advantages / Contributions
借鉴 dropout, 混合max 和 ave 池化,提出 mixed pooling
4 Method
1)mixed pooling 公式
λ \lambda λ is a random value being either 0 or 1
2)mixed pooling 反向传播
先看看 max 和 ave pooling 的反向传播
max pooling
(图片来源于网络,侵删!!!)
ave pooling
(图片来源于网络,侵删!!!)
mixed pooling
得记录下 λ \lambda λ 的取值,才能正确反向传播
the pooling history about the random value λ \lambda λ in Eq. must be recorded during forward propagation.
3)Pooling at Test Time
统计训练时某次 pooling 采用 max 和 ave 的频次 F m a x k F_{max}^{k} Fmaxk 和 F a v e k F_{ave}^{k} Favek,谁的频次高测试的时候该处的 pooling 就用谁,开始玄学了是吧,哈哈哈哈
5 Experiments
5.1 Datasets
- CIFAR-10
- CIFAR-100
- SVHN
5.2 Experimental Results
1)CIFAR-10
train error 高,acc 高
作者解释 This indicates that the proposed mixed pooling outperforms max pooling and average pooling to address the over-fitting problem
可视化结果
可以看出 mixed pooling 包含更多的信息
2)CIFAR-100
3)SVHN
4)Time Performance
6 Conclusion(own) / Future work
LRN
k , n , α , β k, n, \alpha, \beta k,n,α,β 都是超参数, a , b a,b a,b 输入输出特征图, x , y x,y x,y 空间位置, i i i 通道位置
以下内容来自 深度学习的局部响应归一化LRN(Local Response Normalization)理解
import tensorflow as tf
import numpy as np
x = np.array([i for i in range(1,33)]).reshape([2,2,2,4])
y = tf.nn.lrn(input=x,depth_radius=2,bias=0,alpha=1,beta=1)
with tf.Session() as sess:
print(x)
print('#############')
print(y.eval())
LCN
《What is the best multi-stage architecture for object recognition?》
边栏推荐
- 使用Matlab实现:弦截法、二分法、CG法,求零点、解方程
- ERNIE1.0 与 ERNIE2.0 论文解读
- MMDetection安装问题
- Feeling after reading "agile and tidy way: return to origin"
- [introduction to information retrieval] Chapter 3 fault tolerant retrieval
- Implementation of yolov5 single image detection based on pytorch
- 【Programming】
- Use Baidu network disk to upload data to the server
- Implementation of yolov5 single image detection based on onnxruntime
- [CVPR‘22 Oral2] TAN: Temporal Alignment Networks for Long-term Video
猜你喜欢
【Ranking】Pre-trained Language Model based Ranking in Baidu Search
【Random Erasing】《Random Erasing Data Augmentation》
生成模型与判别模型的区别与理解
Agile development of software development pattern (scrum)
Ding Dong, here comes the redis om object mapping framework
SSM supermarket order management system
深度学习分类优化实战
Point cloud data understanding (step 3 of pointnet Implementation)
【信息检索导论】第七章搜索系统中的评分计算
半监督之mixmatch
随机推荐
使用Matlab实现:弦截法、二分法、CG法,求零点、解方程
Calculate the total in the tree structure data in PHP
Get the uppercase initials of Chinese Pinyin in PHP
【MEDICAL】Attend to Medical Ontologies: Content Selection for Clinical Abstractive Summarization
Use Baidu network disk to upload data to the server
[paper introduction] r-drop: regulated dropout for neural networks
MMDetection安装问题
allennlp 中的TypeError: Object of type Tensor is not JSON serializable错误
label propagation 标签传播
【MEDICAL】Attend to Medical Ontologies: Content Selection for Clinical Abstractive Summarization
常见的机器学习相关评价指标
Sparksql data skew
【Torch】解决tensor参数有梯度,weight不更新的若干思路
@Transitional step pit
SSM supermarket order management system
yolov3训练自己的数据集(MMDetection)
Faster-ILOD、maskrcnn_ Benchmark installation process and problems encountered
Drawing mechanism of view (I)
Alpha Beta Pruning in Adversarial Search
Open failed: enoent (no such file or directory) / (operation not permitted)