ConcertoIntraModalObjective<T>

Helpers & Utilities · AiDotNet.ComputerVision.Segmentation.PointCloud

Concerto’s 3D intra-modal self-distillation objective (arXiv:2510.23607), built on the Sonata / DINOv2 online-clustering formulation.

How It Works

A student encoder is optimized to match a momentum-updated teacher across two augmented views of the same point cloud, under a clustering-based cross-entropy. The teacher is never back-propagated into; it only tracks an exponential moving average of the student, which is what keeps the objective from collapsing to a constant.

Two stabilizers from the DINO line are essential rather than decorative:

Methods

Method Summary
T ComputeLoss(Tensor<T> studentLogits, Tensor<T> teacherLogits, Tensor<T> center, double studentTemperature, double teacherTemperature) Cross-entropy between the sharpened, centered teacher assignment and the student’s.
Tensor<T> ComputeTapeLoss(Tensor<T> studentLogits, Tensor<T> teacherLogits, Tensor<T> center, double studentTemperature, double teacherTemperature) The same cross-entropy as ComputeLoss, expressed in engine operations so a GradientTape can differentiate it.
void UpdateCenter(Tensor<T> center, Tensor<T> teacherLogits, double momentum) Advances the running center used to de-bias the teacher’s assignments.
void UpdateTeacher(Vector<T> teacher, Vector<T> student, double momentum) Advances the teacher weights one exponential-moving-average step toward the student.