I'm new to convolutional neural networks and have two related questions:

If all the filters would have the same weights initially, they would all detect the very same feature - so it would be useless to have multiple filters if the weights are not initialized randomly - right?

Is it useful to reset the weights after some time of training of "bad" filters (how to detect such filters?). Maybe due to the initial weights two filters detect (almost) the same feature, and it would be better to let one of those filters re-learn something else?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.