Recurrent processing improves occluded object recognition and gives rise to perceptual hysteresis

Area of Science:

• Computer Vision
• Computational Neuroscience
• Artificial Intelligence

Background:

• Object recognition traditionally modeled as feedforward process, supported by fast human response times and deep feedforward networks.
• Recent shifts suggest recurrent connectivity in the brain significantly contributes to object recognition, particularly for challenging conditions like occlusion.
• Recurrent dynamics may explain perceptual phenomena such as hysteresis.

Purpose of the Study:

• Investigate the benefits of recurrent connections in artificial neural networks for object recognition.
• Systematically compare different recurrent architectures (bottom-up, lateral, top-down) against feedforward models.
• Evaluate the impact of recurrent connections on recognizing partially occluded objects.

Main Methods:

• Introduced three stereoscopic occluded object datasets for evaluating performance.
• Compared parameter-matched recurrent and feedforward artificial neural network architectures.
• Analyzed hidden representations and temporal dynamics of the models.

Main Results:

• Recurrent architectures significantly outperformed parameter-matched feedforward models in occluded object recognition.
• Analysis revealed progressive discounting of occluders over time steps in recurrent models.
• Demonstrated that feedback corrects initial misclassifications and leads to perceptual hysteresis.

Conclusions:

• Recurrent feedback is crucial for robust object recognition, especially under challenging conditions like occlusion.
• Artificial neural networks benefit significantly from incorporating recurrent connections for improved performance.
• The study provides evidence for the role of recurrent dynamics in visual perception and artificial intelligence.