🔗 Source: arXiv

HYPERPARAMETER TRAJECTORY INFERENCE WITH CONDITIONAL LAGRANGIAN OPTIMAL TRANSPORT

🚀 Technical Novelty

  • Mechanism: Learns a data-dependent conditional Lagrangian (kinetic & potential energy terms) to model non-linear hyperparameter-induced dynamics, using optimal transport maps and geodesics to construct a continuous surrogate probability path.
  • Nuance: Differs from standard Euclidean interpolation or conditional flow matching by embedding least-action principles and manifold inductive biases into the cost function, ensuring feasible, physically meaningful trajectories across sparse, high-dimensional hyperparameter spectra.

💡 Yield

  • Empirically outperforms direct interpolation and conditional flow matching baselines in reconstructing conditional probability paths under sparse anchor distributions.
  • Successfully enables inference-time hyperparameter adjustment for reinforcement learning policies (cancer treatment reward balancing) and quantile regression uncertainty bounds without retraining.

⚠️ Limitations

  • Performance degrades with increasing data sparsity, though it degrades less than baselines; requires careful selection of anchor distributions across the hyperparameter spectrum.
  • Currently restricted to single continuous hyperparameters and relies on neural approximations for optimal transport maps, which may face scalability challenges in extremely high-dimensional output spaces.