Overview
Implicit neural representations (INRs) are neural networks that learn to represent a signal (an image, a video, a sound) as a continuous function of coordinates. Instead of storing pixel values directly, the network learns a mapping from coordinates to signal values, essentially compressing the signal into its weights.
This project implements a full meta learning pipeline on top of neural fields: rather than training each network from scratch per signal, a meta learner finds an initialisation that can be quickly adapted to any new signal in just a handful of gradient steps.