EMAN framework grows neural network paths dynamically during training
New EMAN framework lets AI networks sprout new pathways on-the-fly during training
A team of researchers from institutions including the Chinese Academy of Agricultural Sciences and Nanjing University has proposed EMAN (Emergent Modular Atomic Network), a novel framework for multi-task learning that dynamically expands neural network capacity during training. Unlike traditional approaches that rely on predefined architectures or task boundaries, EMAN starts with a single computational path and grows additional independent paths only when persistent optimization evidence emerges.
The framework uses an optimization-driven mechanism to monitor decision signals during training, allowing it to adaptively allocate shared and task-specific representation capacity based on task requirements. In experiments on controlled rank settings, PASCAL-Context, and NYUv2 datasets, EMAN achieved improved performance while maintaining competitive computational costs compared to existing methods.
- EMAN starts with a single neural network path and grows new independent paths dynamically during training when optimization evidence appears
- Achieved improved performance on PASCAL-Context and NYUv2 datasets while maintaining competitive computational costs
- Developed by researchers from institutions including Chinese Academy of Agricultural Sciences and Nanjing University (arXiv:2608.16930)
Why It Matters
This approach could reduce computational waste in AI training while improving multi-task performance through adaptive architecture growth.