Challenging deep learning models with image distortion based on the abutting grating illusion.

Area of Science:

• Computer Vision
• Cognitive Neuroscience
• Machine Learning

Background:

• Current deep learning models lack human-like cognitive abilities.
• Existing image distortions rely on mathematical transformations, not cognitive functions.
• The abutting grating illusion offers a novel approach based on human perception.

Purpose of the Study:

• To introduce a new image distortion method inspired by the abutting grating illusion.
• To evaluate the performance of various deep learning models against this novel distortion.
• To investigate the relationship between model performance and neuroscientific properties.

Main Methods:

• Developed an image distortion technique simulating illusory contour perception using abutting line gratings.
• Tested the distortion on diverse datasets including MNIST and ImageNet silhouettes.
• Evaluated numerous models: trained from scratch, pre-trained on ImageNet, and augmented models.
• Incorporated visualization of early model layers and human subject classification for validation.

Main Results:

• Abutting grating distortion proved challenging for state-of-the-art deep learning models.
• DeepAugment models demonstrated superior performance compared to other pre-trained models.
• Early layer visualizations revealed that better-performing models exhibit the endstopping property, aligning with neuroscience findings.
• Human subject validation confirmed the distortion's effectiveness.

Conclusions:

• The abutting grating illusion provides a robust method for challenging deep learning models.
• Performance on this distortion correlates with neuroscientifically observed properties like endstopping.
• This work highlights the need for developing AI with more human-like perceptual and cognitive capabilities.