YOLO-KAN: Exploring the Adaptability of Kolmogorov-Arnold Networks in the YOLO Model

Traditional Multi-Layer Perceptrons (MLPs) typically employ fixed activation functions and large weight matrices to learn the complex patterns present in these representations. This leads towards large growth in parameters but decreases the interpretability of the model. Kolmogorov-Arnold Networks (KANs) instead employ learnable activation functions inspired by the Kolmogorov-Arnold representation theorem, which is a more flexible way to model nonlinear relationships without inflating the counts of parameters excessively. This study proposes YOLO-KAN, a model based on YOLOv11n, where newly designed KAN modules replace portions of the traditional MLP structures of the backbone. The ablation experiments show that one of the proposed YOLO-KAN models achieves a 1.84% improvement in precision over the yolov11n model in the COCO dataset. These findings suggest that KANs can match or surpass traditional MLPs in computer vision object detection, including large-scale datasets such as COCO. The result can help the community in designing versatile, real-time object detectors for a range of object recognition tasks.