How can max pooling help in making convolutional features as "translational invariant" in a deep learning?



Hi all,

I have read that a CNN model has translation invariance. On exploring more about it, I got to know that the pooling operation in a CNN model does this. Could someone explain what is the logic behind this?


Hi @albela_angur,

Let me break down the answer into two parts.

Translation Invariance vs. Equivariance
First, when the effect of transformation is not detectable by the operator output it is called Translation Invariance. On the other hand, the transformation is deductible then its the case of Translational Equivariant.

Now, the second part. Max pooling achieves invariance in CNN Network. From the behavior of Max pooling function, we can state that the output cannot be retransformed back to the input. Thus, this is Translation Invariance.

The logic behind implementing the Max pooling:

  1. Extracting spatial features
  2. Dimentionality reduction

Hope this answered your query.