Quantifying the cost of network computations to unpack structure-function relationships in the brain

Technical Summary: Quantifying the Cost of Network Computations to Unpack Structure–Function Relationships in the Brain

Problem Statement
While extensive efforts have characterized the anatomical structure of brain networks (from synaptic connectomes to white-matter tracts) and the statistical structure of neural activity, a fundamental gap remains in understanding how network structure directly constrains the computations a network can readily support. Specifically, for a given network structure, which activity patterns are naturally easy to generate, and which are difficult? Addressing this requires a general framework that maps network structure directly onto a quantitative measure of the "computational affordance"—the ease with which any given pattern of activity can be obtained.

Methodology
The authors propose a quantitative framework based on network control theory to define and quantify the cost of computations.

• Definition of Computation: A computation is framed as a goal-directed transition of network activity from an initial state ($x_0$) to a functionally relevant target state ($x_f$), driven by external inputs $u(t)$ over a time horizon $T$.
• Cost Metric: The cost of a computation is quantified as the energy of the input required to implement the transition: $C = \int_0^T \|u(t)\|^2 dt$.
• Linearization and Gramian: To ensure analytical tractability, the network dynamics are linearized around an equilibrium point. The minimum cost to drive the network to a target state is derived as $C_{min} = D^\top W^{-1} D$, where $D$ is the discrepancy between the target and the homogeneous solution, and $W$ is the controllability Gramian.
• Computational Affordance Landscape: The inverse Gramian ($W^{-1}$) is defined as the computational affordance landscape. This landscape encodes the full distribution of costs across all possible activity transformations, shaped jointly by the network's internal interaction structure ($W$) and its input structure ($B$).
• Eigenmode Decomposition: The cost of driving a specific pattern is analyzed by decomposing it into the eigenbasis of the Gramian. The cost is a weighted sum of projections onto eigenmodes, where the weight is determined by the eigenvalues of the landscape.

Key Contributions and Results
The framework is applied across three distinct scales to validate its utility in linking structure to function:

1. Drosophila Head Direction Circuit (Microscale):

   • The framework was applied to a ring-attractor model of the fly head direction circuit.
   • Result: The analysis revealed a single mode in the affordance landscape that was substantially lower in cost than all others. This lowest-cost mode corresponded precisely to the derivative of the activity bump with respect to the heading angle.
   • Implication: The network's energetically cheapest computation is exactly its known biological function: updating the heading representation. Furthermore, the framework predicted the optimal input structure (increasing input ahead of the bump, decreasing behind it), which aligns with known anatomical inputs from shift neurons.

2. Human Brain Connectomes (Macroscale):

   • Structural connectivity data from 100 subjects in the Human Connectome Project (HCP) were analyzed across 7 canonical functional networks.
   • Result: The heterogeneity of the affordance landscape (measured by the interquartile range of eigenvalues) varies systematically along the sensorimotor-association (S-A) axis.
     • Sensory Networks: Display highly heterogeneous landscapes, indicating that some modes are much more expensive than others. This suggests specialization for a specific set of computations (e.g., filtering and refining sensory information).
     • Association Networks: Display more homogeneous landscapes, suggesting uniform access to a broader range of activity states, consistent with flexible, domain-general processing.
   • Spatial Symmetry: The study also found that bilaterally symmetric activity patterns are consistently lower in cost across all networks, aligning with neuroimaging observations of bilateral activation.

3. Artificial Recurrent Neural Networks (RNNs):

   • RNNs trained on decision-making tasks were analyzed to observe how learning reshapes the affordance landscape.
   • Result: Before training, the landscape eigenvalues were tightly clustered (homogeneous). As training progressed, the distribution of costs broadened monotonically, creating distinct low-cost and high-cost modes.
   • Implication: Goal-directed learning actively "sculpts" the affordance landscape, increasing its heterogeneity to favor the specific computations required by the task.

Significance and Claims
The paper establishes a quantitative framework that makes the relationship between network structure and computational function concrete and tractable. By defining a computational affordance landscape, the authors demonstrate that:

• Network structure intrinsically favors certain computations (low cost) while penalizing others (high cost).
• This framework successfully recovers known biological functions (e.g., in the fly compass) and predicts necessary input structures without prior knowledge of the specific task.
• In the human brain, structural differences between sensory and association networks are reflected in the topology of their computational affordances, providing a mechanistic explanation for their distinct functional roles.
• Learning in neural systems can be viewed as the active reshaping of this landscape to optimize for specific goals.

The authors note that while the current implementation relies on linearization around fixed points (which is analytically solvable), the framework is applicable to nonlinear systems by estimating the Jacobian locally. They conclude that this approach offers a principled window into how network structure shapes computational function, with potential applications extending to other biological and physical networks where function emerges from structure.