A Critical Test of Deep Convolutional Neural Networks' Ability to Capture Recurrent Processing in the Brain Using...

Area of Science:

• Computational Neuroscience
• Artificial Intelligence in Vision
• Human Visual Processing

Background:

• Recurrent processing is vital for human vision, aiding tasks like object recognition and segmentation.
• Current deep convolutional neural networks often lack explicit recurrent mechanisms, hindering their ability to model complex visual computations.
• Understanding the neural basis of recurrent processing is key to advancing both neuroscience and AI.

Purpose of the Study:

• To investigate whether deep residual networks (ResNets), despite being feedforward, can capture recurrent processing signals observed in human brain activity.
• To explore the relationship between network depth (modeling varying levels of recurrence) and the ability to explain electroencephalography (EEG) data.
• To differentiate between feedforward and recurrent processing stages in human visual object categorization using computational models.

Main Methods:

• Utilized deep residual networks (ResNets) of varying depths (4 to 34) to simulate excitatory additive recurrence.
• Recorded human electroencephalography (EEG) data during a visual masking paradigm involving an object categorization task.
• Compared the variance in EEG activity explained by ResNets of different depths against human brain responses to masked and unmasked stimuli.

Main Results:

• Deeper ResNets explained significantly more variance in human EEG activity compared to shallower networks.
• All tested ResNets captured differences in brain activity between unmasked and masked trials, emerging around 98 milliseconds post-stimulus onset.
• A notable increase in explained variance by deeper networks was observed after 98 milliseconds, peaking around 200 milliseconds, specifically in unmasked trials, suggesting a role in recurrent processing.

Conclusions:

• Excitatory additive recurrent processing, as approximated by deep ResNets, effectively models aspects of human recurrent visual processing.
• Network depth is a critical factor in capturing the temporal dynamics of neural signals related to recurrent processing.
• The findings provide computational evidence for the timing and nature of recurrent signals in the human visual system.