search

dipamgoswami / ADC

Code for CVPR 2024 paper - Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning
MIT License
12 stars 0 forks source link

Is adversarial images used in distillation loss? #2

Closed YangJae96 closed 1 day ago

YangJae96 commented 4 days ago

Hi.

I have question about the method which I am confused.

In the paper, I can see that in Figure3, the generated adversarial images are fed to old model and estimates logit vector? and used in knowledge distillation loss.

image

But in the above code, I can only see the perturbed images(x, y) is only used in updating the old prototype, which is illustrated in Section 3.3 in your paper, Drift Compensation!

Are the perturbed adversarial images not used with along new task data when training new model?

Thanks in advance.

dipamgoswami commented 4 days ago

Hi,

The adversarial images are used only for the drift compensation (as denoted by the dotted blue lines in Fig. 3 from the paper). And only the new task images are used for the training (as denoted by the red lines in Fig. 3). The perturbed images are not used along with new task data for training.