Path Tortuosity
A metric quantifying the degree of detour in a representation trajectory, defined as the ratio of the trajectory's arc length to the straight-line distance between its endpoints. A larger value indicates a more circuitous path, reflecting the geometric cost of the computational structure itself.
Timeline (last 90 days)
The study finds that when the learned computation requires higher-order interactions among input factors, the representation trajectories systematically become less direct (increased path tortuosity).
The trajectory curvature effect persists under matched performance and controlled training duration, and generalizes from synthetic vector inputs to the training of residual CNNs on rendered images and natural images.
Path tortuosity is the ratio of the arc length of a trajectory to the straight-line distance between its endpoints; values closer to 1 indicate a more direct path, and larger values indicate a more circuitous one.
After controlling for interaction order, path tortuosity can predict the subsequent drift of the representation under continued training and its sensitivity to parameter perturbations, whereas endpoint displacement lacks this predictive ability.
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 developing representation-aware optimization so that it becomes part of the standard training pipeline.
All Facts (5)
The study finds that when the learned computation requires higher-order interactions among input factors, the representation trajectories systematically become less direct (increased path tortuosity).
50%UnverifiedThe trajectory curvature effect persists under matched performance and controlled training duration, and generalizes from synthetic vector inputs to the training of residual CNNs on rendered images and natural images.
50%UnverifiedPath tortuosity is the ratio of the arc length of a trajectory to the straight-line distance between its endpoints; values closer to 1 indicate a more direct path, and larger values indicate a more circuitous one.
50%UnverifiedAfter controlling for interaction order, path tortuosity can predict the subsequent drift of the representation under continued training and its sensitivity to parameter perturbations, whereas endpoint displacement lacks this predictive ability.
50%UnverifiedFuture 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 developing representation-aware optimization so that it becomes part of the standard training pipeline.
50%