But Still It Moves: Static Image Statistics Underlie How We See Motion

Area of Science:

• Computational neuroscience
• Visual perception
• Artificial intelligence

Background:

• Understanding motion perception is challenging due to the interplay of evolution and experience.
• Neural network models offer a way to isolate computational drivers of brain architecture.

Purpose of the Study:

• To investigate the computational basis of motion perception using a neural network (MotionNet).
• To determine if natural image statistics, rather than motion itself, explain key properties of motion processing.

Main Methods:

• Trained MotionNet on moving images to classify velocity.
• Analyzed neural and perceptual biases using the network's architecture and behavior.
• Conducted behavioral tests on humans to assess motion illusions.

Main Results:

• MotionNet replicated biological motion processing properties.
• Neural and perceptual biases for motion direction linked to natural image orientation structure.
• Speed-direction interrelations in MT neurons explained by image autocorrelation.
• Speed-contrast association in natural images identified as key to motion illusions.

Conclusions:

• Natural image statistics are primary drivers of motion perception.
• The speed-contrast association, not the 'slow world' prior, explains motion illusions.
• Integrating artificial systems with biological data advances understanding of complex systems like the brain.