HydraCIL: Decoupled Class-Incremental Learning through Prototype-Guided Multi-Head Classifiers

HydraCIL introduces a decoupled class-incremental learning approach that freezes neural network backbones and uses lightweight task-specific classifiers to enable rapid adaptation on resource-constrained devices. The method achieves competitive performance with state-of-the-art systems while dramatically reducing training time and energy consumption, making it practical for edge AI and embedded applications.

HydraCIL addresses a fundamental tension in machine learning deployment: the gap between computationally expensive training pipelines and the limited resources available in real-world edge devices. Traditional class-incremental learning requires repeated backbone retraining as new tasks arrive, creating bottlenecks for robots, IoT sensors, and embedded systems that must adapt quickly without access to powerful compute infrastructure. By decoupling feature extraction from classification, HydraCIL eliminates this retraining overhead entirely, instead creating lightweight task-specific heads that leverage frozen features extracted once per task.

The broader context reflects growing industry recognition that sustainable AI requires rethinking computational assumptions. As edge deployment accelerates—driven by latency requirements, privacy concerns, and cost constraints—research increasingly prioritizes sample efficiency and energy footprint alongside accuracy metrics. HydraCIL's prototype-guided head selection mechanism provides an elegant solution to the head selection problem, using similarity matching to route inference through the appropriate classifier without complex gating mechanisms.

For developers and organizations deploying AI on constrained hardware, this approach has immediate practical value. Reduced training time translates to faster model iteration cycles, while lower carbon footprint aligns with sustainability commitments. The experimental validation across diverse datasets (CIFAR-100, ImageNet-100, CoRe50, Flowers102) demonstrates generalizability beyond narrow benchmarks.

The significance lies in shifting the continual learning paradigm toward resource-aware design rather than treating efficiency as an afterthought. This research may influence how edge AI architectures evolve, particularly in robotics and autonomous systems where real-time adaptation is essential and retraining windows are measured in minutes, not hours.

• →HydraCIL freezes the backbone network and creates lightweight task-specific classifiers, eliminating expensive retraining cycles
• →The method matches or exceeds state-of-the-art class-incremental learning performance while significantly reducing training time and energy consumption
• →Prototype-guided similarity matching enables efficient head selection at inference without complex gating mechanisms
• →Experimental validation across CIFAR-100, ImageNet-100, CoRe50, and Flowers102 confirms generalizability across diverse datasets
• →The approach makes continual learning practical for embedded systems, robots, and edge devices with strict resource constraints