Shared Visual Abstractions

TL;DR

This paper demonstrates that neural networks and humans share visual representations, as neural networks create abstract art that triggers consistent labels across different architectures and is also recognizable by people.

Contribution

It reveals that neural networks develop shared visual abstractions that are recognizable by humans and trigger consistent labels across various architectures.

Findings

Neural networks produce abstract art that elicits specific labels.

Humans recognize these abstract artworks as representing familiar categories.

Shared visual representations exist across human and machine vision.

Abstract

This paper presents abstract art created by neural networks and broadly recognizable across various computer vision systems. The existence of abstract forms that trigger specific labels independent of neural architecture or training set suggests convolutional neural networks build shared visual representations for the categories they understand. Computer vision classifiers encountering these drawings often respond with strong responses for specific labels - in extreme cases stronger than all examples from the validation set. By surveying human subjects we confirm that these abstract artworks are also broadly recognizable by people, suggesting visual representations triggered by these drawings are shared across human and computer vision systems.