COMP9517 Week8

https://echo360.org.au/lesson/955bbf1e-48dc-4cda-968f-a9dff443e9d5/classroom#sortDirection=desc

https://webcms3.cse.unsw.edu.au/static/uploads/course/COMP9517/20T2/e2306e906b088ac3972b31c8bf1a07e5e9cc48126ffeeaea921030e2cb0f09f5/COMP9517_20T2W8_Deep_Learning.pdf

 

 

1. ML VS DL ：

　　1） 传统方法，我们需要针对图片设计 feature extractor去提取Low/Mid/High - level的features，需要大量的计算

　　2）DL中，自动学习features，do not design features anymore; Images as input, objective functions as Output;

 

 

 

 

2. NN VS CNN：

　　1）NN Weight parameters太多了，CNN是用filter

　　　　例如 Alex-net中 227*227*3 -> 55*55*96  ，共享了11*11*3 * 96 个weights

　　　　　　如果不共享，则55*55*96中的每一个output都要对应一组 11*11*3的weights ； 55*55*96*11*11*3 =? 

　　　　

 

 

　　2） CNN neurons are arranged in 3 dimensions: Width, Height and Depth.

　　2）Neurons in a layer are only connected to a small region of the layer before it (hence not fully connected)