Work

Pre-registration (frozen 7 Nov 2026): replication of the GKL knowledge-distillation gain on CIFAR-100 (arXiv 2305.13948, 2503.08038), as published and against a KD baseline with tuned temperature and loss weights

Year: 2026

Type: article

Abstract: Frozen protocol, decision code, job design, harness, planted-effect table and kernel for a private pre-registered replication, deposited before any confirmatory run. Clause: the mean paired top-1 gain of GKL over KD (ResNet-32x4 teacher, ResNet-8x4 student, three seeds) has a 95% t-interval lower end above 1.75 points (half the published 3.50) with every seed improving, in an as-published arm and in an arm where KD's temperature and loss weights are tuned on a different teacher-student pair. Outcome wording: Replicated / Did not replicate / Inconclusive under protocol X.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.24030585

Author: Goodman, John

Language: English

Topic: Domain Adaptation and Few-Shot Learning

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7225462941

FWCI:

Cites: 2

Cited by:

Related to: 10