Recurrent Processing Dynamics in Occluded Object Recognition Revealed by Electroencephalography and Deep Neural...

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Vision

Background:

• The human visual system effectively recognizes objects despite occlusion.
• The precise temporal dynamics of recurrent processing in occluded object recognition are not well understood.

Purpose of the Study:

• To investigate the temporal dynamics of recurrent processing during occluded object recognition.
• To differentiate feedforward and recurrent contributions using electroencephalography (EEG) and deep neural networks (DNNs).

Main Methods:

• Utilized high-temporal-resolution EEG, backward masking, and DNNs in a two-stage paradigm.
• Applied multivariate pattern analysis (MVPA), temporal generalization analysis (TGA), and representational similarity analysis (RSA).
• Manipulated occlusion levels and employed early masking to disrupt processing stages.

Main Results:

• Low occlusion relied on feedforward processing, while higher occlusion engaged additional processing stages.
• A two-stage recurrent process was identified: an early stage (200-300ms) for low-level features and a late stage (300-500ms) for higher-level representations.
• Early masking disrupted the coordination of these recurrent stages, underscoring their importance.

Conclusions:

• Occluded object recognition involves distinct feedforward and recurrent processing stages.
• Recurrent processing, particularly cross-hierarchical interactions, is essential for robust visual perception of occluded objects.
• The findings clarify the temporal dynamics of recurrence in visual object recognition.