KoPE adds brain-like synchronization to Vision Transformers for faster learning
Neuro-inspired phase encoding boosts training efficiency and cuts data needs by 30%
This paper introduces Kuramoto oscillatory Phase Encoding (KoPE), a neuro-inspired synchronisation mechanism that adds an evolving phase state to Vision Transformers. KoPE improves training, parameter, and data efficiency by accelerating attention concentration. It also enhances performance on semantic and panoptic segmentation, representation alignment with language, and few-shot abstract visual reasoning (ARC-AGI). This approach demonstrates that synchronization can serve as a scalable, neuro-inspired improvement for deep learning models.
- KoPE adds oscillatory phase encoding to Vision Transformers, inspired by neural synchronization in biological brains
- The mechanism improves training and data efficiency by up to 30%, as shown on several vision benchmarks
- Enhances performance on structured tasks like panoptic segmentation and abstract visual reasoning (ARC-AGI)
Why It Matters
Synchronization could make deep models more data-efficient and biologically plausible, reducing compute costs for complex vision tasks.