New framework quantifies how brain networks perform computations
A control theory framework reveals which brain computations cost the least...
Researchers have developed a quantitative framework using control theory to measure how brain network structure shapes the computations a network can support. Their "computational affordance landscape" maps the distribution of costs across all possible activity transitions, revealing which computations a network structure readily supports. Applying it to a circuit model of how insects maintain a sense of direction, they found that updating orientation is the least costly computation, with predicted inputs consistent with known circuitry. In the human brain, the landscape varies by functional role: sensory networks display more heterogeneous landscapes, reflecting specialized information processing, while association networks display more homogeneous landscapes, reflecting generalized processing. In recurrent neural networks trained on cognitive tasks, learning progressively increases landscape heterogeneity, reshaping the distribution of affordable computations. The framework establishes a quantitative approach to structure-function relationships in neural circuits, with future applications extending to other biological and physical networks.
- Introduced a control theory framework quantifying computation costs in neural networks as 'computational affordance landscapes'
- Found sensory networks optimize for specialized tasks while association networks handle generalized processing
- Demonstrated that learning increases landscape heterogeneity in recurrent neural networks
Why It Matters
Enables precise mapping of structure-function relationships in brains and AI networks, potentially advancing both neuroscience and machine learning architectures.