Computer Science 4442A/B Lecture Notes - Lecture 20: Joule, Gradient Descent, Convolutional Neural Network
Document Summary
Traditional object classification training data extract hand- train crafted features classifier new data extract hand- apply crafted features classifier. Why deep networks: hierarchical feature extraction, each stage is a trainable feature transform, the level of abstraction increases up the hierarchy, deep architecture works well for hierarchical feature extraction, hierarchies are used especially natural in vision. , we need : we used to use. , we need d to update w. K kernel n 1 w kj n h k ) output feature map input feature map. Problem with pooling: after several level of pooling, we have lost information about the precise location of things, makes it impossible to use the precise spatial relationships between high-level parts for recognition. ),(1 yx h i yxh i i yxn. )yxn: performed across features and in higher layers, effects, improves variance, improves optimization. Fully connected layer: can have just one fully connected layer, example for 3-class classification problem, eg: 3-class classification problem hn 1.