YOLO-KAN
Researching how Kolmogorov-Arnold Network modules can improve YOLO11n accuracy while reducing network depth.
- best precision
- 65.83%
- peak improvement
- +1.84pp
- reduced layer count
- 299
Project note
This research asks a concrete architecture question: can a Kolmogorov-Arnold Network (KAN) module improve YOLO's feature extraction without making the detector deeper? The study introduces KAN modules into a YOLO11n backbone and evaluates several serialization and flatten-layer strategies on Microsoft COCO.
The strongest configuration reached 65.83% precision, with the largest observed improvement at 1.84 percentage points over the baseline. The experiments also showed that flatten-layer design has a meaningful effect on KAN performance; KAN-2-5 produced the best balance between accuracy and useful feature focus.
The result is not simply a higher score. The simplified architecture reduced the network from 319 to 299 layers, supporting the project's original goal of improving accuracy while reducing depth.
Complete research poster
Architecture, ablations, precision/recall, and heatmaps