Implicit Neural Representation as vectorizer for classification task applied to diverse data structures

Thibault Malherbe

Keywords: INR, SIREN, functa, xgboost

TL;DR: We study usability of implicit neural representation as vectorizer for classification of any type of data (video, image, sound, sensors data)

Abstract:

Implicit neural representations have recently emerged as a promising tool in data science research for their ability to learn complex, high-dimensional functions without requiring explicit equations or hand-crafted features.
Here we aim to use these implicit neural representations weights to represent batch of data and use it to classify these batch based only on these weights, without any feature engineering on the raw data.
In this study, we demonstrate that this method yields very promising results in data classification of several type of data, such as sound, images, videos or human activities, without any prior knowledge in the relative field.

Submission Number: 5