Learning Internal Representations of 3D Transformations From 2D Projected Inputs.

Area of Science:

• Computational neuroscience
• Computer vision
• Mathematical modeling

Background:

• Biological vision systems process 2D projections to perceive 3D structure.
• Understanding the internal representations of 3D transformations is crucial for artificial and biological intelligence.

Purpose of the Study:

• To develop a computational model for inferring 3D structure from 2D motion.
• To investigate how biological vision systems learn 3D transformations from input statistics.
• To demonstrate a proof-of-concept for adaptive internal representations in biological systems.

Main Methods:

• Utilized manifold transport operators to model 3D point transformations.
• Developed a rotational model to infer depth from 2D projected point motion.
• Trained the model using 2D visual stimuli to learn rotational transformations.

Main Results:

• The model successfully learned the generator of the Lie group for 3D transformations from 2D input.
• Demonstrated the model's capability to infer depth from moving 2D points.
• Showcased the model's ability to learn rotational transformations from 2D training data.

Conclusions:

• The computational model provides a framework for understanding 3D structure inference in biological vision.
• The findings suggest that biological systems can adapt internal representations based on sensory input statistics.
• Model performance shows promise when compared to human psychophysical performance in structure-from-motion tasks.