Author: Xie Zhining
Link: https://www.zhihu.com/question/36686900/answer/130890492
Source : Zhihu The
copyright belongs to the author. For commercial reprints, please contact the author for authorization, and for non-commercial reprints, please indicate the source.
Personally, I think there are two main functions:
1. Invariance (invariance), this invariance includes translation (translation), rotation (rotation), scale (scale)
2. Retain the main features while reducing parameters (dimension reduction, the effect is similar PCA) and calculation amount, prevent overfitting and improve model generalization ability
(1) translation invariance:
Here is an intuitive example (digital recognition). Suppose there is a 16x16 picture with a number 1 in it. We need to identify it. This number 1 may be written a little to the left (Figure 1). This The number 1 may be a little to the right (Figure 2). Figures 1 to 2 are equivalent to shifting one unit to the right, but after max pooling, Figures 1 and 2 both become the same 8x8 feature matrix. The main feature we This is captured, while reducing the problem size from 16x16 to 8x8, with translation invariance. The a (or b) in the figure indicates that these a (or b) positions in the original picture will eventually map to the same position.
(2) Rotation invariance:
The following figure shows the recognition of the Chinese character "一". The first one has an inclination angle relative to the x-axis, and the second one is parallel to the x-axis. The two pictures are equivalent to rotation, after multiple max pooling have the same characteristics
(3) scale invariance:
The following figure shows the identification of the number "0". The "0" of the first picture is relatively large, and the "0" of the second picture is smaller, which is equivalent to scaling. Similarly, after many times have the same features after max pooling
The above pictures are all my personal YY, please correct me.