Abstract
In dynamic real-world environments, machine learning models frequently encounter concept drift, leading to performance degradation over time. Traditional mitigation involves "stateless retraining," which is computationally expensive and infrastructure-heavy. This paper proposes Continual-Learning-as-a-Service (CLaaS), a novel framework designed to integrate stateful Continual Learning (CL) strategies into standard MLOps pipelines. Our reference implementation, Continual Brain, demonstrates that stateful updates can achieve up to an 8x improvement in computational efficiency compared to traditional methods while maintaining competitive predictive accuracy.
Introduction
As machine learning moves from static research environments to production, the challenge of non-stationary data becomes paramount. Whether in robotics or e-commerce, the statistical properties of data evolve—a phenomenon known as concept drift. The current industry standard for addressing this is cumulative retraining, where models are periodically retrained from scratch on all available historical data. This approach is inherently inefficient and fails to leverage the stateful nature of learning.
Existing MLOps tools excel at deployment and monitoring but lack native support for stateful updates—the ability to learn from new data without re-processing the entire history or suffering from "catastrophic forgetting."
The CLaaS Framework
The researchers introduce CLaaS as a service-oriented architecture that bridges the gap between academic CL research and industrial application. The framework is designed to be "Edge-Cloud" agnostic, allowing developers to implement complex CL logic without deep domain expertise.
The architecture, as shown in Figure 5, centers on a Service Provider that handles training, model management, evaluation, and logging. It interacts with a dynamic production environment where data arrives as a sequence of "experiences" ($e_0, e_1, ..., e_n$) over time.
Methods and Implementation
The core implementation, Continual Brain, is built upon the Avalanche library and structured into three layers:
- Interface: A REST API and SDK for triggering tasks.
- Middleware: Manages data drift monitoring, versioning, and storage interfacing.
- CL-Backend: Executes strategies like Replay, which maintains a small buffer of past examples to prevent forgetting during updates.
By decoupling the CL logic from the underlying infrastructure (Figure 4), CLaaS allows for seamless integration into existing DataOps and DevOps workflows, providing a standardized way to transition from stateless to stateful model evolution.
Experimental Evaluation
The framework was evaluated using two real-world datasets:
- CORe50: Focused on continuous object recognition for robotics across varying environments.
- DeepFashion-C: Simulating a "New Classes" scenario in a fashion context over 10 successive data batches.
Key Results
- Efficiency: In the CORe50 scenario, CLaaS achieved a 6x reduction in time and memory usage. In the DeepFashion-C scenario, efficiency gains reached nearly 8x.
- Scalability: Unlike cumulative retraining, where time grows linearly with data volume, the Replay strategy maintained stable training times across all experiences.
- Accuracy: Predictive performance remained competitive with the "Cumulative" upper bound, proving that efficiency does not necessitate a loss in model quality.
Conclusion and Future Work
CLaaS represents a significant step toward making Continual Learning a practical reality in production MLOps. By providing stateful updates as a service, it flattens the resource consumption curve associated with model maintenance. Future work will focus on expanding the library of available CL algorithms, exploring memory-free strategies for constrained edge devices, and further automating drift-triggered update cycles.
Summary updated