NORACL: Neurogenesis-Inspired Continual Learning Without Oracle Architecture
A new machine learning method, NORACL, addresses the stability-plasticity dilemma in continual learning by drawing inspiration from biological neurogenesis. The approach eliminates the need for an oracle architecture—a fixed-capacity network sized for an unknown future task stream. Traditional regularization-based methods preserve past knowledge within fixed architectures, which either run out of plastic resources when tasks are weakly related or become over-provisioned when tasks are few or strongly overlapping. NORACL dynamically adjusts representational and plastic resources, allowing the model to remain plastic enough to learn new tasks while staying stable enough to retain previously learned capabilities. The method is detailed in a paper on arXiv (2604.27031), published in April 2025.
Key facts
- NORACL is a continual learning method inspired by neurogenesis.
- It addresses the stability-plasticity dilemma.
- It eliminates the need for an oracle architecture.
- Regularization-based methods rely on fixed-capacity architectures.
- Fixed architectures run out of plastic resources for weakly related tasks.
- Fixed architectures are over-provisioned for few or strongly overlapping tasks.
- The paper is on arXiv with ID 2604.27031.
- The paper was published in April 2025.
Entities
Institutions
- arXiv