ProbXiv
Explorersign in
Problem archiveProblem record

Statement

Let D\mathcal{D} be a natural distribution over examples, let nn be the number of samples, let F\mathcal{F} be a family of interpolating models, and let ε≥0\varepsilon\geq0. A partition LL is an ε\varepsilon-distinguishable feature when it can be learned with test success at least 1−ε1-\varepsilon by training a model from F\mathcal{F} on nn samples labeled by LL. Feature Calibration. For every such D,n,F,ε\mathcal{D},n,\mathcal{F},\varepsilon, and every εL\varepsilon_L-distinguishable feature LL, the distributions of the feature and classifier output and of the feature and true label are statistically close:

(L(x),f(x))f←Train⁡F(Dn)x,y∼D≈ε(L(x),y)x,y∼D.\underset{\substack{f\gets\operatorname{Train}_{\mathcal{F}}(\mathcal{D}^n)\\x,y\sim\mathcal{D}}}{(L(x),f(x))}\approx_{\varepsilon}\underset{x,y\sim\mathcal{D}}{(L(x),y)}.

The conjecture extends feature calibration beyond explicitly supplied labels to any learnable partition, including features unrelated to the class label. The source presents it as a main conjecture, with no resolution supplied.

Record

Comments

Nothing has been published against this problem yet, and nothing has been said about it.

Sign in with an institutional address to take part in the discussion. Reading every thread stays open to everyone.

Sign in

Solve with an agent

Open the statement in a chat, with the problem and the ground rules already written into the prompt.

This opens a third-party site. Nothing is posted back to ProbXiv and nothing you write there is recorded here — what a model gives you is an attempt, which a person still has to check.