Broken time reversal symmetry in visual motion detection.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Animal Behavior

Background:

• Classical models of motion detection assume perfect symmetry in perception when visual stimuli are time-reversed.
• This assumed symmetry is intuitive but may not hold true in biological systems.

Purpose of the Study:

• To investigate whether time reversal symmetry is broken in biological motion perception.
• To identify factors contributing to symmetry breaking in the fruit fly (Drosophila) optomotor response.
• To explore how neural network models of motion detection behave with varying contrast distributions.

Main Methods:

• Designed novel visual stimuli to probe time reversal symmetry in Drosophila's optomotor behavior.
• Trained neural network models to predict scene velocity using natural and artificial contrast distributions.
• Analyzed model responses analytically and numerically to identify sources of symmetry breaking.

Main Results:

• Discovered stimuli that induce broken time reversal symmetry in fruit fly behavior.
• Neural networks trained on naturalistic contrast distributions exhibited broken time reversal symmetry, even with symmetric training data.
• Symmetry breaking in models was linked to contrast asymmetry and other contrast distribution features.
• Shallower neural networks showed stronger symmetry breaking than deeper ones.

Conclusions:

• Biological motion detection systems, like that in Drosophila, often break time reversal symmetry.
• This symmetry breaking may arise from constraints and optimization processes in natural environments.
• Neural network models can replicate and help explain this phenomenon, highlighting the role of data characteristics and model architecture.