arXiv:2406.17814v1 Announce Sort: new
Summary: We study the connection between learnability and strong (or agnostic) learnability for the issue of distribution studying. We present that, opposite to different studying settings (e.g., PAC studying of operate lessons), realizable learnability of a category of likelihood distributions doesn’t suggest its agnostic learnability. We go on to look at what kind of knowledge corruption can disrupt the learnability of a distribution class and what’s such learnability strong in opposition to. We present that realizable learnability of a category of distributions implies its strong learnability with respect to solely additive corruption, however not in opposition to subtractive corruption.
We additionally discover associated implications within the context of compression schemes and differentially personal learnability.
Supply hyperlink