EMAN: New AI Framework for Optimization-Driven Capacity Growth in Multi-Task Learning
Emergent Modular Atomic Network (EMAN), detailed in arXiv:2608.16930v1, introduces an optimization-driven approach to multi-task learning capacity growth. Unlike methods reliant on hard sharing, multiple paths, adaptive sharing, or dynamic expansion—whose capacity changes are constrained by predefined structures or triggered by task boundaries—EMAN starts from single-path computation and grows a new independent path only when persistent optimization evidence emerges. The framework detects an antisymmetric growth direction via latent relative phases without instantiating a second path, and monitors decision signals to turn local optimization evidence into structural decisions. EMAN materializes two equal-capacity paths only after certification, then adaptively allocates shared and task-specific representation capacity.
Key facts
- EMAN is an optimization-driven framework for multi-task learning.
- It starts from exact single-path computation.
- It can grow a new independent path based on persistent optimization evidence.
- It exposes an antisymmetric growth direction through latent relative phases.
- It avoids instantiating a second path until certification.
- It monitors multiple decision signals during training.
- It materializes two equal-capacity independent paths after certification.
- It adaptively allocates shared and task-specific representation capacity.
Entities
—