Regularization on images

I’m working on C1_W4, on this assignments there was no mention of Lambda(regularization term).

Regularization doesn’t help NN improved results in case of images ?

Also, if we don’t convert this images to gray and just use RGB, will it be better than we just trained on(considering the large training time.) ?

Regularization is a more advanced topic that is covered in Course 2 of this series, so please “hold that thought” and stay tuned …

As to whether it works better to train on grayscale images or RGB images, that is “situation dependent”, meaning depends on what your goal is in terms of what you want to detect in the images. If the goal is to detect cats, then grayscale will probably be fine. If the goal is to recognize orange cats, then maybe “not so much” :laughing: … That’s a silly example, but you see the point, right? It depends on what you are trying to discern in the images. Can your human eyes detect the pattern in question in a grayscale image? If so, then that’s probably good enough and will make your training much more efficient.

1 Like

In fact, if you’re interested in pursuing this further, you could actually try some experiments here. Once you finish this assignment, you could take the dataset and convert it to grayscale images. Then try the training again using the 4 layer network that we use in the second assignment here in Week 4. How does the accuracy compare? Of course there are lots of other variables to tweak here, so getting complete and convincing results may not be so straightforward. But it would be pretty easy to just do a basic experiment for curiousity and see if there is any noticeable difference. If you do get motivated to try this, please let us know what you discover one way or the other! You always learn something interesting even if the results turn out not to be as good. That’s useful information as well …

In terms of the idea of experimenting with variations on the test case here in Week 4 Assignment 2, here’s a thread about some experiments with modifying the balance of positive and negative samples in the train and test datasets.