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

HydraCIL is a new decoupled continual learning model that employs prototype-guided multi-head classifiers, designed for efficient deployment in resource-constrained environments. By freezing the backbone and creating lightweight, task-specific classifier heads, HydraCIL avoids extensive retraining, demonstrating competitive performance on CIFAR-100, ImageNet-100, CoRe50, and Flowers102 datasets while significantly reducing training time and carbon footprint. This approach is particularly relevant for practitioners focusing on sustainable AI solutions in real-world applications, such as robotics and edge devices.