Socialized Division and Collaboration (SDC)
What it is
This paper proposes a new framework for class-incremental learning that decomposes learning across multiple specialized models instead of a single monolithic model. It uses an energy-based criterion grounded in Helmholtz free energy to allocate sessions to models based on optimization compatibility, aiming to prevent catastrophic forgetting caused by conflicting optimization directions.
The problem it solves
Catastrophic forgetting and destructive interference in single-model class-incremental learning when successive sessions induce incompatible optimization directions.
What you could use it for
- Class-incremental learning scenarios where new data classes induce optimization conflicts with existing knowledge
- Continual learning systems that suffer from destructive interference when adapting to heterogeneous data streams
- Multi-model collaboration architectures for handling unbounded streams of non-stationary data
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.