Qualitative similarities and differences in visual object representations between brains and deep networks

Area of Science:

• Computer Vision
• Neuroscience
• Cognitive Science

Background:

• Deep neural networks (DNNs) in computer vision show similarities to human visual cortex representations.
• It remains unclear if DNNs replicate qualitative patterns of human perception and brain representations.

Purpose of the Study:

• To investigate whether feedforward DNNs trained for object recognition exhibit known perceptual and neural phenomena.
• To compare DNN representations with human perception and brain activity using distance comparisons.

Main Methods:

• Recasting perceptual and neural phenomena as distance comparisons.
• Analyzing feedforward DNNs, both randomly initialized and trained for object recognition.

Main Results:

• Some phenomena (e.g., global advantage, sparseness) were present in randomly initialized networks.
• Object recognition training induced phenomena like the Thatcher effect and Weber's law.
• Phenomena such as 3D shape processing and surface invariance were absent in trained networks.

Conclusions:

• Findings suggest conditions for phenomena emergence in both brains and DNNs.
• Identifies properties that could enhance DNN capabilities.
• Provides insights into the relationship between artificial and biological vision systems.