CLF: Curve Line Fitting Neural Network Based On Bezier Curve

Jianyi Yang, Guiling Wang

Keywords: CLF, MLP, Interpretable Neural Networks

TL;DR: Propose a network, CLF, directly fits nonlinear relations, eliminating the need for activation functions. It is fully explainable and capable of clearly demonstrating the relationships it learn.

Abstract: The Multilayer Perceptron (MLP) serves as a fundamental architecture in deep learning, leveraging the universal function approximation theorem through linear regression combined with activation functions. Despite its widespread use, the inclusion of activation functions contributes to the inherent nature of MLPs as ``black boxes," limiting their interpretability. In this paper, we propose a novel Curve Line Fitting (CLF) network, which introduces Bezier curve fitting to directly address nonlinear distributions. By replacing traditional linear regression with Bezier curve regression, the CLF network offers a more efficient means of fitting target distributions. Additionally, the removal of activation functions makes the CLF model fully interpretable, enabling clear insights into the relationships between input dimensions and target distributions, as well as the interdependencies across different dimensions. (Sample code for the CLF model will be made available on GitHub.)

Primary Area: interpretability and explainable AI

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10087