TrainingObjectiveKind
Enums · AiDotNet.Enums
Identifies the mathematical objective a trainable model actually optimizes.
Fields
| Field | Summary |
|---|---|
TrainingObjectiveKind ContrastiveAlignment |
A symmetric contrastive (InfoNCE) objective aligns paired inputs from two modalities, as CLIP and CLAP train: the target is the second modality’s batch, paired row by row with the input. |
TrainingObjectiveKind ContrastiveDivergence |
A contrastive-divergence energy objective is optimized. |
TrainingObjectiveKind DiffusionDenoising |
A diffusion denoising or distillation objective is optimized. |
TrainingObjectiveKind EvolutionaryFitness |
Population fitness is optimized by an evolutionary learner. |
TrainingObjectiveKind HamiltonianDynamics |
A scalar Hamiltonian is differentiated with respect to phase-space coordinates and its symplectic derivative field is fitted to observed time derivatives. |
TrainingObjectiveKind LagrangianDynamics |
A scalar Lagrangian is differentiated twice with respect to the state, and the accelerations the Euler-Lagrange equation implies are fitted to observed accelerations. |
TrainingObjectiveKind MultiTask |
Several explicitly structured task objectives are optimized together. |
TrainingObjectiveKind Reconstruction |
The input is reconstructed by an unsupervised learner. |
TrainingObjectiveKind SelfSupervised |
A self-supervised objective defined on the input alone (for example masked prediction of quantized targets among distractors); a supplied target is not used. |
TrainingObjectiveKind Supervised |
A caller-supplied target is optimized by a supervised learner. |
TrainingObjectiveKind VariationalEnergy |
A parameterized quantum-circuit energy is optimized. |
TrainingObjectiveKind ZeroShotEnergy |
A zero-shot partition or energy objective is evaluated without supervised fitting. |
int value__ |