Course 4 - Week 2 - Assignment 2: Why does U-net help prevent overfitting?

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

In section 3 of the assignment, I found this:

Skip connections send information to every upsampling layer in the decoder from the corresponding downsampling layer in the encoder, capturing finer information while also keeping computation low. These help prevent information loss, as well as model overfitting.

I couldn’t recall the last sentence being mentioned in the lecture.

Also the lecture mentioned that in practice, when the number of layers increases, the model performance tends to go down in training. Does it mean overfitting is not a problem for deeper conv net?

2

Does this lecture help?

3

Hi @TRAN_KHANH1

The idea behind that is because 2 things :

  • The skip connnect decrease complexity of the problem by passing information to every upsampling layer in the decoder from the corresponding downsampling layer in the encoder which is carry lower complex information dealing(concatenate) with the downsampling layers(carry high complex information) to balance the complexity of the model
  • The architecture of it looks like the architecture of the Residual Network(Skip connection) if you understand the main concept of it you will find it like that such that the main concept of it is if the model suffer from the overfitting the weight of the skip connection layers is increase and the weights between the skip connection layers tends to zeros

Best Regards,
Abdelrahman

1 Like
5

I watched the video again and still couldn’t find Prof.Andrew mention the overfitting problem. At least, without some reasoning about it.

6

I found the idea quite relatable by now. Thanks for your help!