Memory Replay with Data Compression for Continual Learning (2202.06592v2)

Abstract: Continual learning needs to overcome catastrophic forgetting of the past. Memory replay of representative old training samples has been shown as an effective solution, and achieves the state-of-the-art (SOTA) performance. However, existing work is mainly built on a small memory buffer containing a few original data, which cannot fully characterize the old data distribution. In this work, we propose memory replay with data compression (MRDC) to reduce the storage cost of old training samples and thus increase their amount that can be stored in the memory buffer. Observing that the trade-off between the quality and quantity of compressed data is highly nontrivial for the efficacy of memory replay, we propose a novel method based on determinantal point processes (DPPs) to efficiently determine an appropriate compression quality for currently-arrived training samples. In this way, using a naive data compression algorithm with a properly selected quality can largely boost recent strong baselines by saving more compressed data in a limited storage space. We extensively validate this across several benchmarks of class-incremental learning and in a realistic scenario of object detection for autonomous driving.

• The paper presents MRDC, a novel technique that integrates data compression into memory replay to store more training samples without sacrificing quality.
• It employs Determinantal Point Processes to optimize the trade-off between data quality and quantity, reducing catastrophic forgetting effectively.
• Empirical results on ImageNet and SODA10M benchmarks demonstrate improved accuracy and scalability in real-world continual learning applications.

Analysis of Memory Replay with Data Compression for Continual Learning

The paper "Memory Replay with Data Compression for Continual Learning," authored by Liyuan Wang et al., addresses a critical issue in continual learning – catastrophic forgetting – through a novel approach that balances storage efficiency and information retention by leveraging data compression techniques. This paper extends the established efficiency of memory replay by introducing a method termed Memory Replay with Data Compression (MRDC).

Core Contributions

The MRDC approach innovatively integrates data compression into memory replay strategies to enhance storage capacity without significantly compromising data quality. The method involves compressing past training samples using a parameterizable compression quality, enabling the storage of larger amounts of data within the same memory constraints typically imposed by conventional replay buffers. A significant insight of the research is the nontrivial trade-off between data quality and quantity; compressing data allows for more samples to be retained, which can potentially improve recall and mitigate forgetting, yet excessive compression may distort critical information leading to suboptimal replay outcomes.

To efficiently manage this trade-off, the authors propose employing Determinantal Point Processes (DPPs) to determine optimal compression levels without requiring extensive hyperparameter searches. This approach enables a systematic selection of compression quality that balances the amount and fidelity of stored data, demonstrated empirically across several benchmarks.

Experimental Evaluation

The research extensively validates MRDC across multiple continual learning benchmarks such as class-incremental learning settings on ImageNet-sub and ImageNet-full datasets, and semi-supervised continual learning scenarios in autonomous driving. Notably, on class-incremental scenarios, the implementation of MRDC showcased improvements in averaged incremental accuracies compared to state-of-the-art approaches. For instance, on ImageNet-sub, MRDC bolstered LUCIR's performance by approximately 2.66%, demonstrating robust resistance to forgetting even over numerous incremental phases.

Furthermore, the paper's practical applicability is emphasized through its application in real-world scenarios, notably in semi-supervised object detection tasks using the SODA10M dataset. The ability to compress and effectively replay large-scale object detection datasets suggests significant implications for autonomous systems that must continually integrate new data.

Implications and Future Directions

The paper's findings suggest several implications for future research and development:

1. Scalability and Practicality: MRDC's efficient trade-off handling and compression make it particularly suitable for applications with stringent memory constraints, such as mobile and embedded systems.
2. Integration with Emerging Techniques: As continual learning approaches diversify, integrating MRDC with neural architectures capable of adaptive compression or feature representation learning could amplify its benefits.
3. Real-world Deployment: Augmenting autonomous systems with continual learning capabilities using MRDC can facilitate adaptive intelligence that remains robust over time without excessive computational resources.

In conclusion, the authors present a compelling case for the integration of data compression into memory replay strategies, pointing to enhanced efficiency, scalability, and practical deployment in varied continual learning contexts. This research not only provides an enhancement to existing methodologies but also opens pathways to innovative approaches in handling dynamism within neural network applications.