Imaging an admin that has to handle many new tasks, each an N-way classification problem. Each task has a small amount of S training data (e.g., 20) per class.
The remaining training data are taken by K agents. Each agent only sees C out of the total 100 class. When C is big (e.g., 80), the agents' tasks are similar but slightly differ. When C is small, the agents' tasks should differ more.
For every of its tasks, admin reports accuracy on the standard valid set (excluding unseen classes in that task)
Download cifar100 by
cd images;
./download_data.sh;
All experimental pipeline are included in ./exp.sh. You are suggested to run it code block by code block to get the results in fig.4 to 6.