Does parameter sharing reduces overfitting?

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

When we use parameter sharing, it reduces the number of parameters, shouldn’t it also reduce overfitting compared to the use of all FC nodes?

I think it should but the quiz says otherwise. Why it is wrong?

1 Like
2

Hi rahulkumar,

Good point.

The quiz may say this because In CNNs, it is the pooling layers that specifically serve to control overfitting. Parameter sharing may do so as well, but that depends on how many filters are used. See, e.g.

CS231n Convolutional Neural Networks for Visual Recognition

where it is stated:

“It is common to periodically insert a Pooling layer in-between successive Conv layers in a ConvNet architecture. Its function is to progressively reduce the spatial size of the representation to reduce the amount of parameters and computation in the network, and hence to also control overfitting.”

2 Likes