Always Be Dreaming: A New Approach for Data-Free Class-Incremental Learning

Abstract

Modern computer vision applications suffer from catastrophic forgetting when incrementally learning new concepts over time. The most successful approaches to alleviate this forgetting require extensive replay of previously seen data, which is problematic when memory constraints or data legality concerns exist. In this work, we consider the high-impact problem of Data-Free Class-Incremental Learning (DFCIL), where an incremental learning agent must learn new concepts over time without storing generators or training data from past tasks. One approach for DFCIL is to replay synthetic images produced by inverting a frozen copy of the learner's classification model, but we show this approach fails for common class-incremental benchmarks when using standard distillation strategies. We diagnose the cause of this failure and propose a novel incremental distillation strategy for DFCIL, contributing a modified cross-entropy training and importance-weighted feature distillation, and show that our method results in up to a 25.1% increase in final task accuracy (absolute difference) compared to SOTA DFCIL methods for common class-incremental benchmarks. Our method even outperforms several standard replay based methods which store a coreset of images.

Memory-Free Generative Replay For Class-Incremental Learning

“…Fig. 6 reports the top-1 accuracy of task 2 on old classes (0-19), new classes (20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30)(31)(32)(33)(34)(35)(36)(37)(38)(39), and all seen classes (0-39) with four different ratios. From Fig.…”

“…That means if we only pro-vide a pre-trained model for the starting task, i.e., our problem setting, we would not be able to leverage this strategy. To tackle this issue, [29] and [6] employ synthetic images that generates by data-free knowledge distillation methods [5,32] to alleviate catastrophic forgetting. Comparing with them, although we also utilize such methods, we systematically show that regularization-based methods inherently suffer catastrophic forgetting on fine-grained classes.…”

Absolute Zero-Shot Learning

“…A global model is shared with clients to avoid data leaking. However, data-free knowledge transfer and model inversion techniques (Yin et al 2020;Smith et al 2021;Yin et al 2021) can recover data from the pre-trained model and thus can be used to attack shared models in federated learning. Knowledge distillation utilises the domainexpert teacher model to train a compact student model with pursuing competitive recognition accuracy (Xu, Liu, and Loy 2020;Chen et al 2021).…”

Ex-Model: Continual Learning from a Stream of Trained Models

“…[32] exploits publicly available training data. [3,26] train generative models to extract synthetic samples.…”

“…the model's parameters, that can be used to learn more efficiently. Lastly, (4) is explored in CL with settings such as data-free class-incremental scenarios [3,26], where access to the previous data is forbidden. Again, this scenario assumes a single agent and access to the current data, making it difficult to share knowledge between multiple agents.…”