Recurrent connections facilitate occluded object recognition by explaining-away

Area of Science:

• Neuroscience
• Computer Vision
• Cognitive Science

Background:

• Recurrent connections in the brain are crucial for visual processing but less understood than feedforward connections.
• Object recognition, especially with occlusions, is a key cognitive function potentially relying on recurrent processing.

Purpose of the Study:

• To investigate the role and mechanism of recurrent processing in recognizing occluded objects.
• To demonstrate how recurrent networks can leverage occluder information to improve object recognition.

Main Methods:

• Utilized convolutional models of the visual system, including those approximating the primate visual system.
• Conducted a systematic architecture sweep of convolutional models.
• Performed a human psychophysics experiment to find evidence of the proposed mechanism.

Main Results:

• Recurrent networks exhibit a unique 'explaining-away' computation, unlike feedforward networks.
• This computation uses occluder information to account for missing features, aiding occluded object recognition.
• Evidence consistent with 'explaining-away' was found in both computational models and human experiments.

Conclusions:

• Recurrent connections possess a distinct computational capability to 'explain-away' occlusions, enhancing visual recognition.
• This mechanism offers a potential explanation for how the brain handles partially obscured objects.
• The 'explaining-away' computation may generalize to other perceptual tasks involving context-dependent representation changes.