Local lateral connectivity is sufficient for replicating cortex-like topographical organization in deep neural...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Neurons with similar functions cluster spatially in the primate cortex, forming modular organizations for sensory processing.
• This principle is observed across species, particularly in the visual cortex, where modules are tuned to specific visual features.
• The neural mechanisms driving this topographic organization, especially the role of short-range connections, remain incompletely understood.

Purpose of the Study:

• To investigate whether local lateral connections in artificial neural networks can self-organize into topographic maps resembling the primate visual cortex.
• To explore the computational role of local recurrent structures in developing robust visual representations.

Main Methods:

• Utilized artificial deep neural networks with integrated local lateral connections.
• Trained the networks using standard top-down credit assignment methods.
• Analyzed the emergent topographic organization and compared it to human visual cortex structures.

Main Results:

• Networks spontaneously developed topographic maps analogous to primary, intermediate, and high-level human visual cortex.
• Modular organization emerged without explicit topography-inducing objectives or learning rules.
• Inclusion of local lateral connections improved model robustness against adversarial perturbations.

Conclusions:

• Local lateral connections are sufficient to drive the formation of cortical-like maps in artificial neural networks.
• This self-organization principle suggests a fundamental mechanism for topographic mapping in biological visual systems.
• Local recurrent structures play a crucial role in creating robust visual representations, offering insights into neural computation.