Answers are grounded in this paper and cite their sources.
For a float32 ReLU classifier under a declared binary32 semantics, the classes it assigns to all interventions on a few pixels of a fixed image form a finite object, its finite intervention law; for one target class it is a Boolean function that can be enumerated exactly. We study how a prescribed structure of such a law can be synthesized. On a 784–128–10 MNIST classifier and one anchor image, the target is a predicate on the complete one-pixel and pixel-pair laws: no single-pixel change reaches class t before; afterwards exactly one designated pixel does, above a threshold, and a second pixel blocks it at high values. A closed-form rewrite of one hidden unit, inactive on all 70,000 MNIST images, introduces the predicate for all 9 non-anchor classes and changes no MNIST prediction. Ordinary training with one distinct crafted example, appended to 15–30 declared batches, introduces it for 8 classes, with the blocking pixel and schedule found by exhaustive search and a declared rule. Against a matched control that reproduces the baseline bit for bit, restoring the first-layer weights of either pixel shows that the training change is necessary for both the route and the block; a sham insertion indicates a target-specific response of the dataset for 6 of these classes. For classes 5 and 7, a one-unit edit fitted by a linear program and the trained model have identical laws of the target class on all one-pixel and pair interventions, although their parameter changes differ by orders of magnitude in size and support. All positive results are exhaustive binary32 checks of saved models.
No takes yet. Share an insight, caveat, or question.
Davide Bragetti (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: