Memory-Free Generative Replay For Class-Incremental Learning

Abstract

Regularization-based methods are beneficial to alleviate the catastrophic forgetting problem in class-incremental learning. With the absence of old task images, they often assume that old knowledge is well preserved if the classifier produces similar output on new images. In this paper, we find that their effectiveness largely depends on the nature of old classes: they work well on classes that are easily distinguishable between each other but may fail on more fine-grained ones, e.g., boy and girl. In spirit, such methods project new data onto the feature space spanned by the weight vectors in the fully connected layer, corresponding to old classes. The resulting projections would be similar on fine-grained old classes, and as a consequence the new classifier will gradually lose the discriminative ability on these classes. To address this issue, we propose a memory-free generative replay strategy to preserve the fine-grained old classes characteristics by generating representative old images directly from the old classifier and combined with new data for new classifier training. To solve the homogenization problem of the generated samples, we also propose a diversity loss that maximizes Kullback-Leibler (KL) divergence between generated samples. Our method is best complemented by prior regularizationbased methods proved to be effective for easily distinguishable old classes. We validate the above design and insights on CUB-200-2011, Caltech-101, CIFAR-100 and Tiny Im-ageNet and show that our strategy outperforms existing memory-free methods with a clear margin. Code is available at https://github.com/xmengxin/MFGR.

Incremental Prototype Tuning for Class Incremental Learning

“…Recent popular CIL methods can be categorized into three classes: replay-based, regularization-based, and parameter-isolation-based [4,25,17]. Replay-based methods preserve a small amount of data in previous tasks in memory or a generative model in order to replay these data when training the model on new data to overcome catastrophic forgetting [30,18,39,10,37]. The regularization-based methods provide terms in the final loss to restrict the model with prior or knowledge distillation to change too much to forget previous knowledge [12,14,43,15,9,6].…”

“…Recently, most class incremental approaches fight against catastrophic forgetting by replaying the seen data [30,18,39] or distilling the previous model [15,9,6]. Although distillation and replaying can help to balance plasticity and stability, the representation model learned from the last phase is still needed to be fine-tuned in the next phase endlessly, affecting the whole body of the model, bringing lots of forgetting when just learning a few new concepts, so that limiting the performance of class incremental learning.…”

Few-Shot Class-Incremental Learning via Entropy-Regularized Data-Free Replay

“…[4,20] alternatively uses a GAN architecture to synthesize images, where they fix a trained network as a discriminator and optimize a generator to derive images that can be adopted to distill knowledge from the fixed network to a new network. Recently, [26,30,29] have integrated the idea of generative reply into addressing class-incremental learning problem using Deep-Inversion. However, due to the more challenging setting of FSCIL, these approaches cannot be simply migrated to handle the few-shot scenario.…”