Overview
This paper introduces Online Continual Compression (OCC), a framework designed to learn how to compress data from a non-stationary stream where each sample is seen only once. By utilizing Adaptive Quantization Modules (AQM) based on discrete auto-encoders (VQ-VAEs), the authors successfully mitigate "representation drift" and catastrophic forgetting. This approach allows for significantly more efficient data storage in lifelong learning scenarios without requiring any pre-training.
The Challenge of Online Continual Compression
In the era of big data, systems often face a "firehose" problem: data arrives continuously (e.g., from a self-driving car's sensors), but there is insufficient storage to keep it all in its raw form. While compression is necessary, the data distribution in continual learning changes over time (non-i.i.d.).
Standard compression algorithms like JPEG are static and fail to adapt to specific data distributions. Learned compressors, such as auto-encoders, offer better adaptability but face two massive hurdles in an online setting:
- Representation Drift: If you compress an image today and update your model tomorrow, the new decoder might no longer "understand" the old compressed representation, resulting in corrupted reconstructions.
- Catastrophic Forgetting: As the model learns to compress new types of data (e.g., snowy roads), it might forget how to efficiently compress older, previously seen types (e.g., sunny city streets).
The authors propose a method to learn a compressor on-the-fly that stays consistent and adaptive, overcoming these fundamental challenges.
Core Methodology: Adaptive Quantization Modules (AQM)
The core of the proposed solution is the Adaptive Quantization Module (AQM), which builds upon the Vector Quantized Variational Auto-Encoder (VQ-VAE) framework.
Discreteness as a Stabilizer
The most profound insight of the paper is that quantization acts as a buffer against drift. Instead of storing continuous vectors, VQ-VAEs map inputs to discrete indices in a "codebook."
In a continuous latent space, a small change in model weights shifts every representation. However, in a quantized space, the representation only changes if the encoder's output moves far enough to "jump" to a completely different codebook entry. By stabilizing or slowly updating this codebook, the discrete indices remain a "stable language" between the encoder and decoder over time.
Additional System Components
- Multi-Level Storage (Adaptive Bitrate): AQM uses a sequence of VQ-VAEs representing different compression scales. An incoming sample passes through these levels, and the system selects the highest compression level that still maintains a reconstruction error below a specific threshold. This mimics how humans prioritize memory—keeping clear details of novel/complex items while highly compressing familiar ones.
- Self-Replay: To prevent catastrophic forgetting, the model reconstructs random samples from its own memory and uses them to perform gradient updates. This keeps the model familiar with past data distributions.
- Decoupled Training: Modules are trained greedily and independently. This prevents gradients from later modules from interfering with earlier ones, avoiding the instability of end-to-end backpropagation in non-stationary online environments.
Key Results across Diverse Domains
The authors evaluated AQM across several domains, proving its versatility and effectiveness:
- Continual Learning (CIFAR-10): On disjoint CIFAR-10 benchmarks, AQM achieved 43.5% accuracy with a small memory buffer, significantly outperforming standard Experience Replay (27.5%) and JPEG-based compression (33.9%).
- High-Dimensional Sensor Data (Kitti LiDAR): AQM achieved a 32x compression rate on LiDAR data while maintaining a reconstruction error of only 18.8 cm—accurate enough for downstream SLAM (Simultaneous Localization and Mapping) applications.
- Reinforcement Learning (Atari): In environments like Pong and Ms. Pacman, AQM compressed states by 16x while maintaining nearly identical performance on linear probe tasks, proving that the compression preserves the critical "meaning" of the game states.
Limitations and Future Work
While AQM provides a robust foundation for online continual compression, the authors note a few areas for future exploration:
- Temporal Correlation: The current model treats every frame as independent. In video or RL, exploiting the massive redundancy between consecutive frames could yield even higher compression rates.
- Intelligent Prioritization: The model currently relies on a variation of Reservoir Sampling. Future iterations could explore deciding which samples are most worth keeping based on their "surprise" factor or utility to the downstream learner.
- Computational Overhead: Running multiple VQ-VAE levels and self-replay adds computational cost, which might be a constraint for deployment on very low-power edge devices.
Summary updated