Beyond scaling: how brains reorganize to support higher intelligence

Area of Science:

• Systems neuroscience
• Computational neuroscience
• Artificial intelligence

Background:

• Intelligence is traditionally viewed as a whole-brain property.
• Previous theories emphasized feedback-driven computation.

Purpose of the Study:

• To formalize the concept of intelligence as a whole-brain property using Reinforcement Learning and Approximate Dynamic Programming (RLADP).
• To propose four distinct levels of vertebrate intelligence (rodent, primate, human, cetacean) based on feedback signal processing architectures.
• To explore the relationship between neural architecture, energy strategies, and intelligence levels.

Main Methods:

• Formalization using Reinforcement Learning and Approximate Dynamic Programming (RLADP).
• Analysis of conserved allometric rules for cortical ion channels across mammalian species.
• Investigation of white matter as an active communication system and the role of the corticothalamic loop.

Main Results:

• Vertebrate intelligence is categorized into four distinct architectural levels.
• Human neurons uniquely deviate from ion channel scaling rules, reallocating energy to white matter.
• White matter's superlinear scaling presents a communication-energy cost trap.
• The corticothalamic loop is critical for timing forward-backward cortical processing; timing degradation leads to intelligence failure.

Conclusions:

• Intelligence evolution is driven by architectural shifts in feedback processing and energy allocation, not just parametric changes.
• Biological strategies for managing communication-energy trade-offs in the brain offer insights into artificial intelligence limitations.
• Understanding these biological constraints is crucial for advancing both neuroscience and artificial intelligence.