Predictive coding explains asymmetric connectivity in the brain: A neural network study.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Visual Perception

Background:

• Predictive coding frameworks propose hierarchical generative models with feedback and forward pathways.
• Symmetric feedback connections aid visual processing of noisy stimuli in computational models.
• Neurophysiological studies suggest asymmetric cortical feedback connections exist in the brain.

Purpose of the Study:

• Investigate the role of neural feedback in computing grasp parameters using artificial neural networks.
• Evaluate the impact of asymmetric feedback, inspired by cortical structures, on visual processing.
• Determine the optimal configuration of predictive feedback for robust grasp detection.

Main Methods:

• Utilized convolutional neural network models augmented with predictive feedback.
• Trained models to compute grasp positions for real-world objects.
• Systematically tested asymmetric feedback connections of varying ranges and origins/terminations.

Main Results:

• Symmetric feedback improved grasp detection on noisy stimuli.
• Medium-range asymmetric feedback demonstrated optimal performance enhancement under adverse conditions.
• The most significant improvements occurred when feedback originated from layers closer to the input.

Conclusions:

• Biologically realistic asymmetric predictive feedback enhances neural network robustness to noisy visual stimuli.
• Medium-range asymmetric feedback, originating from proximal layers, is particularly effective for grasp detection.
• This study provides insights into the neural mechanisms underlying visual processing and object interaction.