I am trying to generate a class-incremental framework which decreases forgetting on a split-cifar10 dataset but I am running into problems.
I am trying to use a memory buffer and train my model on experience replay. However, I am getting a dip in the middle of my model. Accuracies are as follows: [62.8, 27.3, 25.7, 28.499999999999996, 84.2].
I have tried to increase the complexity of the buffer and make the experience replay balance the replay buffer but it is not improving.
