Research map / Optimization / Implicit bias and generalization
Tractable Sharpness-Aware Learning of Probabilistic Circuits
Hrithik Suresh, Sahil Sidheekh, Vishnu Shreeram M P, Sriraam Natarajan, Narayanan Chatapuram Krishnan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Hessian-trace regularizer (gradient-based) | Einsum Network · Synthetic manifolds, 1% data | test NLL reduction (%): 49.53 mean reduction; 65.71 DoF; 89.24 sharpness | Improved |
| add Hessian-trace regularizer (gradient-based) | Einsum Network · Synthetic manifolds, 100% data | test NLL reduction (%): 5.05 mean reduction; 0.49 DoF; 35.25 sharpness | Improved |
| add Hessian-trace regularizer (gradient-based) | Einsum Network · Synthetic manifolds, twisted_eight 100% data | test NLL reduction (%): -1.22 | Hurt |
| add Hessian-trace regularized EM | PyJuice HCLT · Binary density, 1% data | test NLL reduction (%): 7.12 mean; 8.81 DoF; 23.04 sharpness | Improved |
| add Hessian-trace regularized EM | PyJuice HCLT · Binary density, 5% data | test NLL reduction (%): 2.19 mean; 5.79 DoF; 15.32 sharpness | Improved |
| add Hessian-trace regularized EM | PyJuice HCLT · Binary density, 50% data | test NLL reduction (%): -0.48 NLL; 0.21 DoF; 5.55 sharpness | Hurt |
| add Hessian-trace regularized EM | PyJuice HCLT · Binary density, 100% data | test NLL reduction (%): -0.31 NLL; 0.06 DoF; 4.04 sharpness | Hurt |
| add Hessian-trace regularized EM | PyJuice HCLT · Binary density, dna 1% data | test NLL reduction (%): 29.45 | Improved |
| increase regularization strength mu beyond 0.1 (ablation) | PyJuice HCLT · DNA 5% split | validation NLL: small mu in (0, 0.1] captures most gains; larger mu marginal gains at cost of slight underfitting | Mixed |
| use sum-squared-gradient trace instead of autograd Hessian trace | Einsum Network · 2D spiral, network depth scaling | computation time: autograd runtime exponential blow up vs linear scaling for SSG | Improved |
Closest papers by meaning
- Scaling Probabilistic Circuits via Monarch Matrices (ICML 2025 · similarity 0.90)
- On the Hardness of Approximating Distributions with Tractable Probabilistic Models (NeurIPS 2025 · similarity 0.89)
- The Expressivity Boundary of Probabilistic Circuits: A Comparison with Large Language Models (NeurIPS 2026 · similarity 0.88)
- A Universal Class of Sharpness-Aware Minimization Algorithms (ICML 2024 · similarity 0.87)
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.87)
- Scaling Tractable Probabilistic Circuits: A Systems Perspective (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
