Abstract:
This study examines the scope of Kolmogorov–Arnold Networks (KAN) for image classification in low-resource computational environments. Conventional deep learning models for vision tasks require significant hardware resources, often limiting their use in academic, rural, or resource-constrained settings. The research explores whether a KAN- based architecture can achieve competitive accuracy while operating on a CPU-only system under strict runtime limits. The proposed architecture integrates two KANConv2d layers, each leveraging spline-based functional approximations derived from the Kolmogorov–Arnold representation theorem, followed by pooling, normalization, and a spline-driven classifier head (KANLinear2). This design aims to balance expressive power with computational efficiency. The model was evaluated on eight datasets representing diverse domains: MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, STL-10, EuroSAT- RGB, and two medical imaging benchmarks from the MedMNIST suite (ChestMNIST and PathMNIST). All experiments adhered to a fixed 30-minute training budget on an Intel Core i3 CPU with 8 GB RAM. Results demonstrate that the KAN-based model achieves 99.1% accuracy on MNIST and 91.0% on Fashion-MNIST, while CIFAR-10 reaches 72.3% and CIFAR-100 achieves 43.5% under identical resource constraints. Medical datasets exhibit promising performance, with PathMNIST reaching 84.1% and ChestMNIST achieving 66.0%. Complexity analysis indicates a parameter range of 0.10– 0.13 million across tasks, with per-image compute ranging from 20.8 MFLOPs to 302 MFLOPs, primarily driven by input resolution. These outcomes confirm the suitability of KAN architectures for constrained environments without reliance on GPUs. The findings highlight the trade-off between accuracy and efficiency while presenting a reproducible phybrid architectures, lightweight optimization techniques, and domain-specific adaptations to enhance performance under the same resource envelope. This work establishes a practical foundation for deploying advanced learning methods in settings where computational power is limited.
