# Implicit Neural Representations with Periodic Activation Functions

**Authors:** Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, Gordon Wetzstein  
**arXiv:** [2006.09661](https://arxiv.org/abs/2006.09661)  
**Date:** 2020-06-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.09661v1/

## Key points

- Authors introduce Sirens as a novel approach to signal representation using neural networks
- Sirens address limitations of current network architectures in capturing fine details and spatial/temporal derivatives of signals defined implicitly by partial differential equations
- Sirens use periodic activation functions for representing complex natural signals and their derivatives effectively
- Authors develop principled initialization scheme based on Siren activation statistics
- Sirens can effectively represent various types of signals such as images, wavefields, video, sound, and their derivatives
- Sirens can be utilized to solve challenging boundary value problems including Eikonal equations, Poisson equation, Helmholtz equation, and wave equations
- Authors explore combining Sirens with hypernetworks to learn priors over the space of Siren functions
- This innovative approach enhances signal representation through implicit neural networks with periodic activation functions

## Summary

In their paper titled "Implicit Neural Representations with Periodic Activation Functions," authors Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein introduce a novel approach to signal representation using neural networks. The authors address the limitations of current network architectures in capturing fine details and spatial/temporal derivatives of signals defined implicitly by partial differential equations by proposing the use of periodic activation functions in their sinusoidal representation networks (Sirens). Through their research, they demonstrate that Sirens are well-suited for representing complex natural signals and their derivatives. By analyzing Siren activation statistics, the authors develop a principled initialization scheme and showcase the effective representation of various types of signals such as images, wavefields, video, sound, and their derivatives. Furthermore, they show how Sirens can be utilized to solve challenging boundary value problems including Eikonal equations (yielding signed distance functions), the Poisson equation, and the Helmholtz and wave equations. Additionally, the authors explore combining Sirens with hypernetworks to learn priors over the space of Siren functions. This innovative approach opens up new possibilities for enhancing signal representation through implicit neural networks with periodic activation functions. Overall, this paper offers valuable insights into advancing signal processing techniques and solving complex physical problems efficiently using neural network-based representations.