Balancing Selectivity and Generality in Object Recognition through Structured Interconnectivity.

Balancing selectivity and generality in object recognition is a significant challenge, as it requires the ability to discern fine details that set objects apart while simultaneously embracing the common threads that classify them into one single category. Here we investigated how the brain addresses this challenge by examining the relationship between the interconnectivity of neural networks, the dimensionality of neural space, and the balance of selectivity and generality using neurophysiological data and computational modeling. We found that higher interconnectivity in the TEa of macaques IT cortex was associated with lower dimensionality and increased generality, while lower interconnectivity in the TEO correlated with higher dimensionality and enhanced selectivity. To establish the causal link, we developed a brain-inspired computational model formed through Hebbian and anti-Hebbian rules, with wiring length constraints derived from biological brains. The resulting structured interconnectivity created an optimal dimensionality of the neural space, allowing for efficient energy distribution across the representational manifold embedded in the neural space to balance selectivity and generality. Interestingly, this structured interconnectivity placed the network in a critical state that balances adaptability and stability, and fostered a cognitive module with cognitive impenetrability. In summary, our study underscores the importance of structured interconnectivity in achieving a balance between selectivity and generality, providing a unifying view of balancing two extreme demands in object recognition.