Category-specific perceptual learning of robust object recognition modelled using deep neural networks

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Machine Learning

Background:

• Real-world object recognition faces significant ambiguity.
• The human visual system exhibits remarkable robustness to noisy conditions.
• Perceptual learning is a potential mechanism for acquiring visual robustness.

Purpose of the Study:

• Investigate the role of perceptual learning in visual robustness for humans and deep neural networks (DNNs).
• Determine if training with noisy object images yields category-specific or general robustness improvements.
• Assess DNNs as models for human perceptual learning.

Main Methods:

• Trained humans and DNNs with object images in Gaussian noise.
• Evaluated noise thresholds for accurate recognition before and after training.
• Conducted layer-wise analysis of DNNs to understand learning mechanisms.

Main Results:

• Humans showed category-specific robustness improvements after training.
• Standard DNNs exhibited both category-general and category-specific learning.
• DNNs pre-trained to match human accuracy showed only category-specific learning, mirroring human results.
• Layer analysis indicated category-general learning in lower DNN layers and category-specific learning in higher layers.

Conclusions:

• Visual robustness to noise is acquired through learning.
• Humans likely develop robustness from everyday exposure to visual noise.
• Category-specific improvements in robustness involve higher-level visual representations in both humans and DNNs.