Data Science Stack Exchange is a question and answer site for Data science professionals, Machine Learning specialists, and those interested in learning more about the field. It only takes a minute to sign up.
Sign up to join this community
I originally trained multiple individual binary classifiers for each label of an image. Then, I realized I can train a single multilabel model for this task. I used binary_cross_entropy loss for this instead of categorical_cross_entropy, but besides for changing the loss function I made no major changes. However, I find that the multilabel classifier still substantially underperforms the individual label classifiers. Is this common and to be expected? Are there any tricks I am missing?
Thanks!
It is expected for at least two reasons:
The multilabel classifier can use the extra information about other classes eg. by inferring "I want to find horses. I also know houses and dogs. While horses/houses are easy to divide, I have to be careful about the division of horses and dogs.". More concrete: Think of it as transfer learning. Eg. when using any kind of neural network, early feature-extraction layers profit from the extra loss and are learned more generally when you train a single model compared to many binary models.
Classes in your multilabel problem are guaranteed to be more balanced than classes in your binary problems are.
