Abstract
Incremental learning (IL) aims to acquire new knowledge from current tasks while retaining knowledge learned from previous tasks. Replay-based IL methods store a set of exemplars from previous tasks in a buffer and replay them when learning new tasks. However, there is usually a size-limited buffer that cannot store adequate real exemplars to retain the knowledge of previous tasks. In contrast, data distillation (DD) can reduce the exemplar buffer's size, by condensing a large real dataset into a much smaller set of more information-compact synthetic exemplars. Nevertheless, DD's performance gain on IL quickly vanishes as the number of synthetic exemplars grows. To overcome the weaknesses of real-data and synthetic-data buffers, we instead optimize a hybrid memory including both types of data. Specifically, we propose an innovative modification to DD that distills synthetic data from a sliding window of checkpoints in history (rather than checkpoints on multiple training trajectories). Conditioned on the synthetic data, we then optimize the selection of real exemplars to provide complementary improvement to the DD objective. The optimized hybrid memory combines the strengths of synthetic and real exemplars, effectively mitigating catastrophic forgetting in Class IL (CIL) when the buffer size for exemplars is limited. Notably, our method can be seamlessly integrated into most existing replay-based CIL models. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing replay-based baselines.
Introduction
Class Incremental Learning (CIL) presents a significant challenge: models must continuously learn new classes without suffering from "catastrophic forgetting" of previously acquired knowledge. The standard approach involves a replay buffer that stores a small subset of old data (exemplars). However, researchers face a fundamental trade-off between data quality and quantity. Real data is high-quality but space-intensive, limiting the diversity of the buffer. Conversely, Data Distillation (DD) creates information-dense synthetic images, but its effectiveness plateaus as the buffer size increases.
As illustrated in the figure below, synthetic memory initially outperforms real memory at very low capacities, but real memory eventually provides better performance as more storage becomes available. The proposed hybrid memory consistently outperforms both across all buffer sizes.
Proposed Methodology
The core of this work is the Hybrid Memory buffer, which integrates both synthetic and real data to leverage their complementary strengths. The framework introduces two primary innovations:
Continual Data Distillation (CDD)
Traditional data distillation is computationally prohibitive for incremental tasks because it often requires training numerous models from scratch to identify representative data. CDD addresses this by distilling synthetic data from a sliding window of checkpoints captured during the current task's training. This approach is not only more efficient but also ensures the distilled data is specifically tailored to the model's actual learning trajectory.
Conditional Real Data Selection
Rather than selecting real exemplars through random sampling or standard "herding" techniques, the authors employ a greedy algorithm. This selection process is conditioned on the existing synthetic data, aiming to pick real samples that provide the most complementary information. By minimizing the distance between the hybrid buffer's distribution and the full original dataset, the method ensures that limited buffer space is used optimally.
Experimental Results
The hybrid memory approach was evaluated on standard CIL benchmarks, including CIFAR-100 and TinyImageNet. The method was integrated into several state-of-the-art CIL baselines, such as iCaRL, BEEF, and FOSTER.
Experimental results consistently show that synthetic and real data are complementary. Synthetic data captures the "essence" or mean features of a class, while real data captures the nuances and edge cases that distillation misses. Through grid search, the authors found that a 0.5 ratio (half synthetic, half real) typically yields the best performance. As shown below, the hybrid memory approach consistently improves the Average Incremental Accuracy (AIA) across all tested baselines.
Conclusion
This paper demonstrates that a hybrid approach to memory replay effectively addresses the limitations of both real and synthetic data in Class Incremental Learning. By optimizing a hybrid buffer through Continual Data Distillation and conditional selection, the proposed method achieves superior performance in memory-constrained scenarios. The seamless integration with existing baselines and robust performance across benchmarks highlight its practical utility for future CIL research.
Summary updated