Machine learning is more framework. The knowledge is more important. Like, why does batch normalization cause problem ? how do you increase model performance without change model architecture? Which optimizer gives better performance SGD vs Adam ?