The idea in plain language.
Estimate score behavior along sampled directions instead of computing a full expensive derivative structure.
How it works
Conventional score matching can require operations involving the trace of a Hessian. Sliced score matching replaces that computation with an objective based on random projections, using Hessian-vector products that automatic differentiation can compute efficiently. The paper analyzes the estimator and evaluates it in density estimation and related generative-learning tasks.
What to keep in mind
Efficiency and estimation quality depend on the model and projection procedure. Theoretical guarantees rely on the assumptions stated in the paper.
Source: Sliced Score Matching: A Scalable Approach to Density and Score Estimation. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.