WHY CNN?
If use DNN fully connect network, when there are lots of features especailly in image processing, over thousands or millions dimensions in a vetor, it will become a very complicated issue.
We can filter some dimensions by leveraging prior knowledge that decreasing the parameters in DNN. It’s CNN, which is much easier than DNN.
- A neural doesn’t have to see the whole pecture to discover the pattern
- The same pattern appears in the different part of the image. e.g. one in upper-left and another in the middle of the picture.
- Subsampling the pixels to make the image smaller and do not influence the pattern detection.