Download presentation
Presentation is loading. Please wait.
1
Convolutional Neural Network
2
Outline 1.What is a convolutional neural network? 2.How does it works ? 3.The derivation of CNN. 4.CNN in caffe
3
1.What is a convolutional neural network?
Filters 特点： 权值共享、局部连接、降维。降低了网络模型的复杂度、减少了权值的数量
4
2.How does CNN works ? 2.1 Convolutional layer(卷积层) 1 3 2 5 6 9 8 1 4
3 2 5 6 9 8 ×1 ×0 ×1 1 4 ×0 ×1 ×0 ⊗ = ×1 ×0 ×1 Image Filter Convolved feature
5
2.How does CNN works ? 2.2 Pooling layer(池化层) 8 Max pooling 1 3 4 8 2
4 8 2 5 7 6 9 4 Avg pooling Convolved feature Pooled feature
6
3.Derivation of CNN 3.1 FL BP算法的权值更新 (1) Forward: (2)
激活函数：sigmoid, tanh, Relu… Goal: Minimize E
7
3.Derivation of CNN a、FL BP算法的权值更新 (3) (4) (5) (6)
是第l层第i个结点的残差等于第l+1层与其连接的所有结点的权值和残差的加权和再乘以该点对z的导数值 参数(第l层第j个结点到第l+1层第i个结点的权值)的导数等于第l层第j个结点的激励值乘以第l+1层第i个结点的残差 (5) (6)
8
4.在 Caffe 中如何计算卷积？ Feature vector: C×K×K Image: C×H×W
More feature vectors… Image: C×H×W
9
× … … 4.在 Caffe 中如何计算卷积？ … Image: C×H×W Feature Matrix: (H×W) ×(C×K×K)
Filters: Cout×C×K×K Filter Matrix: Cout ×(C×K×K)
Similar presentations
