Structured interconnectivity optimizes neural geometry for balancing specificity and generalization in object...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• Object recognition requires balancing specific details with general categories.
• Neural network structure is key to how biological and artificial systems achieve this balance.
• Understanding this balance is crucial for advancing both neuroscience and artificial intelligence.

Purpose of the Study:

• Investigate the relationship between neural network interconnectivity, neural space dimensionality, and object representation abstraction.
• Examine how the brain balances specificity and generalization in object recognition.
• Establish a causal link between neural structure and recognition capabilities.

Main Methods:

• Combined neurophysiological data from macaque monkeys with computational modeling.
• Analyzed interconnectivity in areas TEa and TEO of the inferior temporal (IT) cortex.
• Developed a brain-inspired computational model constrained by empirical wiring length.

Main Results:

• Higher interconnectivity in area TEa correlated with lower dimensionality and greater generalization.
• Lower interconnectivity in area TEO correlated with higher dimensionality and greater specificity.
• The computational model demonstrated that structured interconnectivity creates optimal dimensionality for efficient energy distribution, balancing specificity and generalization.

Conclusions:

• Structured neural connectivity is critical for robust object recognition.
• Multi-level abstraction in object representation is enabled by optimized neural space dimensionality.
• Findings provide insights into the neural mechanisms underlying visual perception and inform AI development.