Transform-invariant recognition by association in a recurrent network

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Object recognition is invariant to changes in viewpoint, scale, and retinal position.
• A memory trace, linking consecutive views of an object, is hypothesized to underlie this invariance.
• Sustained neural activity and recurrent connectivity are potential mechanisms for this memory trace.

Purpose of the Study:

• To investigate if extensive recurrent connectivity in the visual pathway can provide the memory trace necessary for view-invariant object recognition.
• To model how synaptic efficacies, based on associations between object views, contribute to sustained neural activity.
• To analytically explore the network's behavior under varying association strengths.

Main Methods:

• Development of a simplified analytical model of the visual pathway with high recurrent connectivity.
• Incorporation of synaptic efficacies derived from associations between pairs of object views.
• Analysis of network dynamics as the relative strength of association between views increases.

Main Results:

• A critical threshold in the association strength between views was identified.
• Below the threshold, sustained activity is driven by individual views.
• Above the threshold, the network consistently converges to a stable object representation, irrespective of the initial stimulus view.

Conclusions:

• Extensive recurrent connectivity can generate a memory trace supporting view-invariant object recognition.
• A specific regime of association strength enables the network to reliably represent objects.
• The model demonstrates the capacity to store a large repertoire of objects, each defined by a finite set of views.