Deep Learning
- CNNs have won several competitions
-
ImageNet, Kaggle Facial Expression, Kaggle Multimodal Learning, German Traffic Signs, Handwriting, ….
-
CNNs are deployed in many practical applications
-
Image recognition, speech recognition, Google’s and Baidu’s photo taggers
-
xCNNs are applicable to array data where nearby values are correlated
-
Images, sound, time-frequency representations, video, volumetric images, RGB-Depth images, …
-
CNN is one of the few deep models that can be trained in supervised way.
- Easy to understand and implement.
What is the neural ?¶
Multilayer neural network¶
- And return to repeat step 1-3 until error is smaller than threshold
Back propagation¶
https://towardsdatascience.com/understanding-backpropagation-algorithm-7bb3aa2f95fd
http://galaxy.agh.edu.pl/%7Evlsi/AI/backp_t_en/backprop.html
- \(\eta\) 学习率
学习率低，收敛慢，容易掉到坑里陷入局部最优；学习率高，可能使得收敛过程不稳定，来回震荡，一直不收敛
idea:
- 设置不同的学习率，看哪种情况最好
- 设计一个自适应学习率。此时学习率不再固定，可以通过外在条件算(梯度，学习要有多快，特征权重的大小...)
Deep Learning For Image Understanding¶
Advantages for convolution¶
Example:
- 200x200 image
- 10 filters of size 10x10
- 10 feature maps of size 200x200
- 400,000 hidden units with 10x10
- fields=1000 parameters
为什么引入CNN可以大规模减少权数参数训练量因为CNN通过
1） 局部连接（Local Connectivity）
2） 权值共享（Shared Weights）
3） 池化（Pooling）
- 来降低参数量
Pooling¶
CNN池化可以通过池化层来降低卷积层输出的特征维度，在有效减少网络参数的同时还可以防止过拟合现象
https://zhuanlan.zhihu.com/p/78760534
CNN in Detail¶
-
步长？
-
Refer to the PPT !
- 传统图像分类：分段；深度学习图像分类：端到端
- 交叉商
- Minimize Loss - Gradient Descent.
创建日期: 2024年1月5日 23:32:29