# Polynomial Implicit Neural Representations For Large Diverse Datasets

**Authors:** Rajhans Singh, Ankita Shukla, Pavan Turaga  
**arXiv:** [2303.11424](https://arxiv.org/abs/2303.11424)  
**Date:** 2023-03-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.11424v1/

## Key points

- The paper introduces a novel approach to implicit neural representations (INR) for signal and image representation
- Authors propose representing images with polynomial functions instead of positional encodings to enhance representational power
- Element-wise multiplications between features and affine-transformed coordinate locations after every ReLU layer enable a more powerful representation of large and diverse datasets
- The Poly-INR model developed through this approach performs comparably to state-of-the-art generative models without convolution, normalization, or self-attention layers while using significantly fewer trainable parameters
- The research has been accepted at CVPR 2023, indicating its significance in the computer vision community
- Code for implementing the Poly-INR model is available on GitHub at https://github.com/Rajhans0/Poly_INR

## Summary

The paper titled "Polynomial Implicit Neural Representations For Large Diverse Datasets" by authors Rajhans Singh, Ankita Shukla, and Pavan Turaga from Arizona State University introduces a novel approach to implicit neural representations (INR) for signal and image representation. <br>
INR has gained popularity in various end-tasks such as superresolution and 3D modeling. However, existing architectures often rely on sinusoidal positional encoding which limits their representational power due to finite encoding sizes. To address this limitation, the authors propose representing images with polynomial functions instead of positional encodings. This allows for a higher degree of polynomial representation without the need for additional encoding layers.<br>
The proposed method involves element-wise multiplications between features and affine-transformed coordinate locations after every ReLU layer. This enables a more powerful representation of large and diverse datasets. The Poly-INR model developed through this approach is evaluated qualitatively and quantitatively on datasets like ImageNet.<br>
Remarkably, the Poly-INR model performs comparably to state-of-the-art generative models without convolution, normalization, or self-attention layers while using significantly fewer trainable parameters. This breakthrough in INR modeling not only enhances representational power but also paves the way for broader adoption of these models in complex generative modeling tasks.<br>
Furthermore, the paper has been accepted at CVPR 2023, underscoring the significance of this research in the computer vision community. The code for implementing the proposed Poly-INR model is available on GitHub at https://github.com/Rajhans0/Poly_INR.<br>
Overall, this work presents a promising advancement in implicit neural representations that holds great potential for applications across various domains requiring sophisticated signal and image processing techniques.