PFSKANs: A Novel Pixel-Level Feature Selection Model Based on Kolmogorov-Arnold Networks

Modern computer vision relies heavily on deep learning architectures to process visual data. Prior research has shown that Convolutional Neural Networks (CNNs) and transformers dominate the landscape of image classification and feature extraction. These traditional models often function as black boxes, making it difficult to understand which specific input components drive final predictions. While attention mechanisms provide some insight, they do not offer direct interpretability at the individual pixel level during the initial selection phase. Existing feature selection methods frequently struggle with high-dimensional image data or require complex iterative processes to identify relevant regions. The reliance on fixed kernel sizes in CNNs can also limit the flexibility of the model to adapt to non-grid-like feature importance. This absence of evidence motivated the development of a more transparent alternative that leverages the inherent interpretability of spline-based neural architectures.

This research introduces a novel framework called PFSKANs to perform direct pixel-level feature selection within image classification tasks. The investigators sought to create a fundamentally distinct alternative to standard trainable Convolutional Neural Networks (CNNs) and vision transformers. By modifying Kolmogorov-Arnold Networks (KANs), the team aimed to detect key pixels with high contribution scores directly at the input stage. The study focuses on visualizing the selection procedure to ensure that the model remains interpretable for human observers. Another primary objective involved developing a mathematical approach to identify and dimensionally standardize these interpretable pixels for subsequent processing. The project evaluates whether this new architecture can match the performance of established models while offering superior transparency. Researchers specifically targeted the MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets to validate the model's versatility across different levels of image complexity.

The researchers developed the Pixel-level Feature Selection (PFS) model by adapting the simplification techniques found in Kolmogorov-Arnold Networks (KANs). This architecture utilizes trainable selection procedures that allow for the intuitive visualization of pixel importance across various datasets. The experimental framework applied this model to four benchmark image classification datasets: MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100. Unlike iterative methods, the proposed selection procedure is performed only once to identify the most significant input features. A specific mathematical approach was implemented to handle the dimensional standardization of the identified interpretable pixels. The team compared the PFSKANs against standard CNN baselines to assess their relative merits. Performance metrics focused on three specific areas: classification accuracy, parameter efficiency, and total training time relative to established convolutional models.

Experiments on the MNIST and Fashion-MNIST datasets demonstrate that PFSKANs achieve accuracy levels comparable to traditional Convolutional Neural Networks (CNNs). The model successfully detected key pixels with high contribution scores directly at the input image level across all tested datasets. Testing on the more complex CIFAR-10 and CIFAR-100 datasets confirmed that the architecture maintains high performance even with increased visual variance. Data regarding parameter efficiency showed that the KAN-based approach requires a competitive number of trainable weights relative to established vision models. Training time measurements indicated that the single-pass selection procedure does not impose significant computational overhead compared to standard training loops. The visualization of the selection process confirmed that the model identifies interpretable features that align with the structural components of the input images. These results indicate that the PFSKANs model provides a robust alternative for feature selection without sacrificing the speed or accuracy expected of modern computer vision systems.

The study establishes PFSKANs as a viable and interpretable alternative to black-box architectures in computer vision. These findings suggest that Kolmogorov-Arnold Networks (KANs) can be effectively adapted for high-dimensional image classification tasks. The ability to perform pixel-level feature selection in a single trainable step offers a new path for efficient model design. Future research may apply this mathematical approach to other domains where input interpretability is a primary requirement. The researchers conclude that the proposed model bridges the gap between high-performance deep learning and transparent feature selection. This work provides a foundation for developing more efficient and understandable vision systems for industry and academic applications. By demonstrating comparable accuracy and parameter efficiency to CNNs, the authors highlight the potential for KAN-based models to redefine standard practices in image processing.