Statistical inference · Embeddings & representations · Optimization & optimizers / Open question
Can Neural Networks Recover Interaction Graphs?
Neural networks can represent complex interactions parsimoniously. Turning their learned features into a trustworthy scientific interaction graph requires more than a good fit.
Seed catalog · Showing the bundled record.
Problem statement
For binary graphical models with interaction order at most , at most factors per spin and , fit centered local energies with width- neural interaction-screening estimators. Establish finite-sample hypergraph recovery by a conditional-influence threshold, explicitly connecting neural approximation error, achieved optimization error and minimum influence. Determine when compact neural structure yields better cost than enumerating all degree- monomials; representation alone is not an optimizer guarantee.
Assumptions and scope
- Fix the discrete alphabet, interaction family, identifiability/minimum-influence condition and a bounded neural hypothesis class; unrestricted compact neural representations need not be efficiently learnable.
- Specify the centered local-energy parameterization, regularization, initialization and optimizer. Convexity in the function being represented does not imply convexity in network weights.
- Define structure through conditional dependence or another invariant criterion. A threshold on raw first-layer weights requires its own justification under rescaling and cancellation; independently learned conditionals may also require a compatibility check.
What would settle it?
A finite-sample support-recovery theorem with explicit approximation/optimization/threshold errors for a nontrivial model family, or an obstruction showing which additional structure or algorithmic changes are necessary.
Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.
Status and evidence
The NeurISE paper explicitly asks for theoretical guarantees on learned structure. GRISE already provides guarantees in a fixed linear basis; NeurISE supplies function-space/global-optimum arguments and empirical neural recovery. Its 2026 discrete-diffusion application is not a checked resolution of finite-training graph recovery. The target is the neural structural guarantee, not the invention of interaction screening. This is an editorial reformulation of a published open problem. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.
Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.
Sources and further reading
- Abhijith J., Lokhov, Misra and Vuffray — Learning of Discrete Graphical Models with Neural Networks ↗problem statement · source checked 2026-11-19
PDF Sections 3–5 and p.10 checked: NeurISE uses a nonconvex neural parameterization of a function-space screening objective; structure guarantees and removal of data biases are explicitly posed as unanswered questions. Infinite-data global-optimum arguments are not finite-SGD support-recovery guarantees.
- Vuffray, Misra and Lokhov — Efficient Learning of Discrete Graphical Models ↗partial theory · source checked 2026-11-19
Primary abstract checked: GRISE generalizes screening to discrete alphabets, arbitrary bases and multi-body factors under an identifiability condition. High-order model representation can still have a large computational cost.
- Elamvazhuthi, Jayakumar and Lokhov — Discrete Diffusion with Sample-Efficient Estimators for Conditionals ↗related algorithm · source checked 2026-11-19
Primary abstract checked: round-robin discrete diffusion using NeurISE conditionals, evaluated on Ising/Potts, images and scientific data. This is a later neural-screening application, not an abstract-level resolution of finite-training interaction-graph recovery.