Decoding self-motion from visual image sequence predicts distinctive features of reflexive motor responses to visual...

Area of Science:

• Neuroscience
• Computer Vision
• Human Perception

Background:

• Visual motion analysis is vital for detecting objects and self-motion, aiding action planning.
• Understanding how the brain processes visual motion for motor responses is key to human-environment interaction.

Purpose of the Study:

• To investigate if convolutional neural network (CNN)-based image motion analysis for self-motion decoding mirrors human reflexive responses.
• To compare the spatiotemporal frequency tuning of CNN decoders with human ocular and manual responses to visual motion.

Main Methods:

• Training a CNN to decode self-motion from human natural movements.
• Analyzing the spatiotemporal frequency tuning of the CNN decoder.
• Comparing decoder tuning with reflexive ocular/manual responses and visual stimulus properties.
• Manipulating training data to observe effects on tuning specificity.
• Examining representational differences for head-axis rotation versus transversal motion.

Main Results:

• CNN decoder tuning peaked at high-temporal and low-spatial frequencies, matching reflexive responses but differing from visual image properties.
• Training data manipulations significantly altered decoder tuning specificity.
• Differences in tuning for center-masked stimuli between rotational and transversal motion mirrored discrepancies in ocular and manual responses.
• Head-axis rotation was decoded via spatial accumulation, while transversal motion involved complex spatial interactions.

Conclusions:

• The study supports the hypothesis that visual motion analysis for reflexive motor responses is acquired for decoding self-motion.
• CNN models provide insights into the mechanisms underlying visual self-motion perception and motor control.
• Findings highlight the shared principles between artificial and biological visual motion processing for interaction with the environment.