Machine Learning 0006 How much labeled data is needed for training
I didn't know how much training data was needed before, I thought the more the better. I read an article recently, which is not the conclusion of my own experiments. I think this conclusion is more reliable.
1. In general, given 5000 labeled samples per class , a model with acceptable performance can be trained.
2. With at least 10 million labeled data, a model that is close to or surpasses the human level can be trained.