The Geometry of Representation Formation: Bures Trajectories Reveal a New Dimension of Neural Network Learning

From Endpoint to Path: Why the Process of Representation Formation Deserves Study

Traditionally, our analysis of neural network representations has focused almost entirely on the endpoint state after training is complete—examining what features the model has learned and how the representation space is organized. However, how the internal structure forms step by step—the process itself—is rarely measured systematically.

A paper titled "The Information Geometry of Representation Formation" proposes a new perspective: treating representation learning as a geometric trajectory, using Bures geometry to measure movement within representation space and quantum Fisher information to characterize local sensitivity to input factors. According to this research, the information contained in the training path may far exceed what the endpoint reveals.

The core insight of this approach is: if the formation trajectory of representations has systematic geometric features, these features may tell us more about the learning process itself—not just what the model ultimately "knows," but also how it "came to know" it.

Bures geometry is a natural Riemannian metric defined on the space of quantum states (density matrices), induced by the Bures distance. On classical probability distributions, it is closely related to the Fisher information metric; in the quantum case, it measures the minimum statistical distinguishability between two quantum states. When neural network representations are treated as probability distributions or covariance structures, Bures geometry provides an intrinsic, parameterization-insensitive measure, giving the question "how far has the representation moved in representation space" a rigorous mathematical definition without relying on extrinsic coordinate choices such as Euclidean distance.

Quantum Fisher Information (QFI) is the generalization of classical Fisher information in quantum statistics. It characterizes the sensitivity of a quantum state to an infinitesimal change in some parameter, and its reciprocal gives the optimal precision lower bound (the quantum Cramér-Rao bound) for unbiased estimation of that parameter. In representation analysis, treating input factors as parameters to be estimated and the network's internal representation as a quantum state, QFI can quantify the local sensitivity of the representation to each input dimension, yielding a sensitivity profile independent of the specific network architecture.

Core Findings: Sensitivity Concentration and Goal-Driven Geometric Reorganization

The study first reveals an intuitive but previously unquantified geometric phenomenon: learning concentrates local sensitivity on task-relevant factors. Measured with quantum Fisher information, the representation's sensitivity to various input factors is redistributed during training—task-relevant dimensions gain higher sensitivity while irrelevant dimensions are compressed.

More strikingly, when only the prediction target is changed, the representation geometry reorganizes around the corresponding visual factors. This means the geometric structure is not fixed but flexibly reflects the demands of the current computational task—change the objective function, and the geometry reshapes accordingly.

This finding elevates representation geometry from a static descriptive tool to a dynamic diagnostic tool: by observing how geometry reorganizes with task changes, we can infer what kind of computation the model is constructing.

Representation reorganization refers to the phenomenon where, when the task objective changes, the structure of the network's internal representation space—not just the output—undergoes systematic adjustment. Unlike the slow drift caused by continued training, reorganization typically manifests as a rapid reversal in the ranking of dimension importance: previously compressed dimensions regain sensitivity, and previously dominant dimensions are suppressed. This concept differs from "catastrophic forgetting," which emphasizes the loss of performance on old tasks, whereas reorganization focuses on the reconfiguration of geometric structure, which can occur even when old and new tasks have highly overlapping input distributions. Observing reorganization requires measurement tools that can align representation spaces across training stages—precisely the advantage of intrinsic metrics like Bures geometry over neuron-by-neuron comparison.

Path Tortuosity: The Geometric Imprint of Higher-Order Interactions

Perhaps the most counterintuitive finding of the study is: when the learned computation requires higher-order interactions among input factors, the representation trajectory systematically becomes less direct.

In plain terms, if a task requires combining multiple input dimensions (e.g., XOR-like problems requiring the interaction of two features), the path the representation travels during training becomes more curved and circuitous. This "path tortuosity" becomes a geometric indicator of computational complexity.

Crucially, this effect persists under matched performance and controlled training duration, and generalizes from synthetic vector inputs to the training of residual CNNs on rendered and natural images. This means path tortuosity is not a byproduct of insufficient training or performance differences, but a geometric manifestation of the computational structure itself.

Higher-order interactions in function learning refer to outputs that depend on the joint values of two or more input dimensions and cannot be decomposed into the sum of individual contributions from each dimension. XOR is the simplest example of a second-order interaction: no single dimension carries predictive information; only their combination is meaningful. Higher-order interactions (e.g., third-order, fourth-order) correspond to more complex compositional structures. In neural networks, higher-order interactions typically require multiple layers of nonlinear transformations to be constructed step by step, thus lengthening the "detour distance" of the representation from its initial state to a task-adapted state. Path tortuosity quantifies this detour: it is the ratio of the trajectory's arc length to the straight-line distance between its endpoints; values closer to 1 indicate a more direct path, and larger values indicate more circuitousness. Since this metric controls for training duration and final performance, it can relatively purely reflect the geometric cost of the computational structure itself.

Formation History Predicts Representation Stability

If tortuosity were merely a marker of computational complexity, its value would remain limited. But the study found a more practically significant property: formation history contains predictive information about representation stability.

After controlling for interaction order, path tortuosity predicts the subsequent drift of the representation under continued training, as well as its sensitivity to parameter perturbations—whereas endpoint displacement lacks this predictive ability. In other words, how a representation "came to be" says more about how stable it is than where it "ultimately is."

This has direct implications for model interpretability and robustness research: if we care whether a representation will drift on new data or be sensitive to parameter perturbations, tracking its formation trajectory is more predictive than examining its endpoint state.

Actionable Geometry: Actively Shaping Training Trajectories

The research goes beyond measurement. The authors demonstrate that representation-aware optimization methods can actively shape these trajectories, showing that formation geometry is not only measurable but also actionable.

This opens new possibilities for training interventions: if we want more stable representations, we might introduce path constraints into the optimization objective; if we want more flexible representations, there may be corresponding geometric guidance strategies.

Significance and Outlook

This work establishes representation formation as a new source of information: its geometry reveals the properties of the computation being constructed, predicts subsequent representation stability, and provides targets for intervention during optimization. For researchers concerned with interpretability, robustness, and training dynamics, it offers a practical framework situated between "black-box behavior analysis" and "purely theoretical analysis."

Future directions may include: extending Bures geometric trajectory analysis to larger-scale models and more complex task structures; exploring the feasibility of path tortuosity as an early-training diagnostic indicator; and further developing representation-aware optimization to become part of standard training pipelines.

One-line Summary

This research treats representation learning as Bures geometric trajectories, finding that trajectory tortuosity predicts representation stability and can be actively optimized.

Paragraph Summary

This study proposes analyzing the formation process of neural network representations from an information geometry perspective, rather than focusing only on the training endpoint. It uses Bures geometry to measure movement in representation space and quantum Fisher information to characterize local sensitivity to input factors. Core findings include: learning concentrates sensitivity on task-relevant factors, and changing the prediction target causes geometry to reorganize around corresponding factors; when tasks require higher-order interactions among input factors, representation trajectories systematically become more curved, even under matched performance and controlled training duration; path tortuosity also predicts representation drift under continued training and sensitivity to parameter perturbations, which endpoint displacement cannot do. The authors further show that representation-aware optimization can actively shape training trajectories, making formation geometry both measurable and actionable.