The paper "Improved Implicit Neural Representation with Fourier Reparameterized Training" addresses the low-frequency bias issue in multi-layer perceptrons (MLPs) used in Implicit Neural Representations (INRs). The authors propose a Fourier reparameterization method that learns a coefficient matrix of fixed Fourier bases to compose the weights of MLPs. This approach aims to alleviate the spectral bias, which often leads to poor approximation accuracy for high-frequency components in signals like complex textures and intricate geometric shapes. The key contributions of the paper are: 1. **Theoretical Analysis**: The authors theoretically prove that appropriate reparameterization can reduce the spectral bias by altering the magnitude of gradients from different frequencies. 2. **Fourier Reparameterization Method**: They propose a practical reparameterization method for MLPs, which does not modify the network architecture but improves approximation accuracy. 3. **Experimental Validation**: The method is evaluated on various INR tasks, including 1D function approximation and real-world vision applications, showing superior performance in terms of convergence speed and approximation accuracy. The paper also includes detailed experimental results and ablation studies to validate the effectiveness of the proposed method. The authors conclude that their Fourier reparameterization method can improve the representation accuracy for a wide range of commonly used INR network architectures, providing a new approach to adjust the learning bias of neural networks.The paper "Improved Implicit Neural Representation with Fourier Reparameterized Training" addresses the low-frequency bias issue in multi-layer perceptrons (MLPs) used in Implicit Neural Representations (INRs). The authors propose a Fourier reparameterization method that learns a coefficient matrix of fixed Fourier bases to compose the weights of MLPs. This approach aims to alleviate the spectral bias, which often leads to poor approximation accuracy for high-frequency components in signals like complex textures and intricate geometric shapes. The key contributions of the paper are: 1. **Theoretical Analysis**: The authors theoretically prove that appropriate reparameterization can reduce the spectral bias by altering the magnitude of gradients from different frequencies. 2. **Fourier Reparameterization Method**: They propose a practical reparameterization method for MLPs, which does not modify the network architecture but improves approximation accuracy. 3. **Experimental Validation**: The method is evaluated on various INR tasks, including 1D function approximation and real-world vision applications, showing superior performance in terms of convergence speed and approximation accuracy. The paper also includes detailed experimental results and ablation studies to validate the effectiveness of the proposed method. The authors conclude that their Fourier reparameterization method can improve the representation accuracy for a wide range of commonly used INR network architectures, providing a new approach to adjust the learning bias of neural networks.