This paper provides a fair and comprehensive comparison between Kolmogorov–Arnold Networks (KAN) and Multi-Layer Perceptrons (MLP) across various tasks, including machine learning, computer vision, natural language processing, audio processing, and symbolic formula representation. The authors control the number of parameters and FLOPs to ensure a fair comparison. Their main findings are: 1. **Performance Comparison**: - **Symbolic Formula Representation**: KAN outperforms MLP, primarily due to its B-spline activation function. - **Other Tasks**: MLP generally outperforms KAN in machine learning, computer vision, NLP, and audio processing. 2. **Ablation Studies**: - Replacing MLP's activation function with B-spline significantly improves its performance in symbolic formula representation tasks, making it comparable to KAN. - The B-spline activation function is the primary factor contributing to KAN's advantage in symbolic formula representation. 3. **Continual Learning**: - KAN exhibits more severe forgetting issues compared to MLP in a standard class-incremental continual learning setting. 4. **Conclusion**: - The paper highlights that while KAN has unique advantages in symbolic formula representation, MLP remains superior in other tasks. The differences in performance are largely attributed to the activation functions used in KAN and MLP. The authors aim to provide insights for future research on KAN and other MLP alternatives, emphasizing the importance of understanding the functional differences between these models.This paper provides a fair and comprehensive comparison between Kolmogorov–Arnold Networks (KAN) and Multi-Layer Perceptrons (MLP) across various tasks, including machine learning, computer vision, natural language processing, audio processing, and symbolic formula representation. The authors control the number of parameters and FLOPs to ensure a fair comparison. Their main findings are: 1. **Performance Comparison**: - **Symbolic Formula Representation**: KAN outperforms MLP, primarily due to its B-spline activation function. - **Other Tasks**: MLP generally outperforms KAN in machine learning, computer vision, NLP, and audio processing. 2. **Ablation Studies**: - Replacing MLP's activation function with B-spline significantly improves its performance in symbolic formula representation tasks, making it comparable to KAN. - The B-spline activation function is the primary factor contributing to KAN's advantage in symbolic formula representation. 3. **Continual Learning**: - KAN exhibits more severe forgetting issues compared to MLP in a standard class-incremental continual learning setting. 4. **Conclusion**: - The paper highlights that while KAN has unique advantages in symbolic formula representation, MLP remains superior in other tasks. The differences in performance are largely attributed to the activation functions used in KAN and MLP. The authors aim to provide insights for future research on KAN and other MLP alternatives, emphasizing the importance of understanding the functional differences between these models.