Issue with EWC Plugin on SMNIST Benchmark #1662
Unanswered
rch-huang
asked this question in
Feature Request
Replies: 1 comment
-
|
This is the expected behavior, EWC does not work in class incremental continual learning. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
🐛 Describe the bug
When using the EWCPlugin with non-multi-head models, such as MLP, on the SMNIST benchmark, the accuracy of learned experiences like Exp000, Exp001 are almost 0.00 or very low, with only the current experience being okay. This issue does not occur with the PMNIST benchmark or when using a multi-head MLP within the SMNIST benchmark.
When using the official example provided in the repository ewc_mnist.py, this problem persists.
BTW, I am curious why the GEM plugin works with MLP models on the SMNIST benchmark. Could this be because EWC inherently supports domain incremental learning (e.g., Permuted MNIST) better than class incremental learning (e.g., Split MNIST)?
🐜 To Reproduce
python3 ewc_mnist.py --scenario=smnist --ewc_mode=separate
🐞 Screenshots
'''
Beta Was this translation helpful? Give feedback.
All reactions