In lectures, we know that the error analysis could help us find what specific problems decrese the accuracy on earth. But I wonder after this how should we do to figure the problem out? E.g., if we find the dog mistake is the main problem in recognizing cats, how can we reslove the dog mistake?
If still use the ways of extend the network, use a bigger network, then what’s the meaning of using error analysis to find the dog mistake? If we don’t know the low accuray is mainly caused by dog mistake, we also use that method to decrease the error. What I mean is extending the network is a general method but not solely specific for dog mistake. Thank you everyone.
Prof Ng discusses this general topic quite a bit in the lectures on Error Analysis. It may be worth just watching those lectures again with your question in mind and listening again carefully to what he says. You could think of all of Course 3 as about different variations of this type of question. What to do when your solution isn’t working as well as you need it to. What you do next depends on what you find. If the problem is that your network frequently mistakes dogs for cats, then the next question is how many “negative” training samples with pictures of dogs do you have? If that number is low, then that would suggest you need more training data of dog images with appropriate labels. But if your error rate is also higher than you want in terms of false negatives on cats, then maybe you have a high bias problem and need a deeper/bigger network. It all depends on the details of what you find.
1 Like
Thank you so much sir. The latter part of “If that number is low, then that would suggest you need more training data of dog images with appropriate labels. But if your error rate is also higher than you want in terms of false negatives on cats, then maybe you have a high bias problem and need a deeper/bigger network.” dispels my doubt. Actually, there are and only are 2 situations for dog mistakes. One is the network didn’t learn to judge the dog to negative for there is not enough dog samples to learn. The other is the network can not recognize dog to negtive while we have enough dog samples. This is caused by the ability of network, so I need to enlarge my network to let it distinguish the subtler difference between dogs and cats. Is there any mistakes in my comprehension?
Great! The way you explained the different cases makes sense to me. Onward! 