Questions about the accuracy of the base task

[ShiWuxuan]

Hi Chunwei,

Your work is wonderful and has been very rewarding for me. But I have a slight doubt as to how your method achieves amazing accuracy when learning the underlying task on three datasets. In particular, on ImageNet-Subset, the accuracy of the initial 50 classes can be close to 100% (from Figure 7 it is estimated to be 98%).

Look forward to and thank you for your reply.

[horsepurve]

Thanks for the question!

The accuracy on the base classes is also higher on CIFAR-100 (5/10/20 phases, Figure 5). This is because:

1. Training on the base classes is an easier task compared to training on the full classes. There are only ~half of the total classes to distinguish.
2. Other methods all simply used the out-of-the-box classification result from the network, but here ours involves many "post-processings" like Voronoi subdivision, test-time assignment, etc.