INFERNO: Inference-Aware Neural Optimisation

de Castro & Dorigo (2019) — INFERNO

Use

practical adjacent categorization method; not exact hard D partition theory.

DS18 audit (31 Aug 2026)

registered a bibliography key here because this is the closest applied statement of the DS18 objective. INFERNO's loss is the inverse Hessian of an Asimov binned-Poisson likelihood — a profiled-Fisher surrogate — but the feasible set is a soft, differentiable histogram of a neural summary. What transfers: the objective family and the confirmation that nuisance-aware binning is the target practitioners want. What does not: hard partitions, exactness, uniqueness, strict isolation in decision distance, empirical-to-population transfer, and finite one-point exchange stability — INFERNO reports no optimality theorem at all.

Idea

differentiable optimization of binned summaries against an inference objective.