C4_W1 Quiz Clarification of Explaination

Hello, I’m not understanding the explaination provided here:

If “transfer learning is not bound to ConvNets”, then why would this not be a valid/applicable attribute?


Let me give you an example:

YOLO was trained on many different classes, human, animals, cars, etc. So, we can use that model on a different data of human, animals, cars, etc. Here data is different but task is same: classifying human, animals, cars, etc.

But can we apply that model for a different task, for example, complex brain images or geological structure? Here, data and task both are different, so, we cannot do transfer learning.

This is what I think a possible reason why this choice is wrong. Open to be rectified…


Yes, this issue has come up before. Perhaps the wording could be a little clearer. My understanding is that what they meant by the question is “Which of the following are uniquely true about convolutional layers?” In other words, what they are getting at is how are convnets different than fully connected nets. Since this is our first introduction to convnets here in week 1, a lot of the material is about why convnets are interesting and useful.

So the point is that yes, you can do transfer learning with convnets, but you can also do transfer learning with FC nets and RNNs and so forth. So that is not something that is special about convolutions.

I think that quiz needs an update, the question and the explanation don’t seem to support each other.

1 Like