GMvandeVen
commented
1 year ago Thank you for your kind message. The way you describe the difference between the task- and class-incremental experiments in Table 2 & 3 seems correct to me. With class-incremental learning the model has to distinguish between all possible classes, while with task-incremental learning the model only has to distinguish between the classes within the task.
I'm not completely sure what kind of references you are requesting, but I think this paper by Zenke et al. (2017, ICML) might be useful: http://proceedings.mlr.press/v70/zenke17a. In Figure 3 of this paper they perform Split MNIST in a way that in our paper we call task-incremental.
I hope this helps, but let me know if it is not clear yet.
Young-Jo-Choi
I recently read your paper titled 'Three types of incremental learning', and I'm very grateful for the meticulous breakdown. In table2,3, you presented evaluations for the same 'method' under both task-incremental and class-incremental learning scenarios. The only difference between two scenarios is whether giving or not giving 'context' information on training and evaluating processes, isn't it?
It's relatively easy to understand the class-IL evaluations since there are many papers on class-IL methods, so I assume it's about calculating accuracy across all labels. For the task-IL evaluations, I assume it's about calculating accuracy by limiting the predicted space within the labels of a given context, and then averaging this for all contexts. Is this how you did it, or did you use a different approach?
If you have any references or other papers that you referred to for your task_IL accuracy measurement, I would appreciate it if you could share them with me.