Pre-registration protocol: Independent replication of Müller, Kornblith & Hinton (2019), "When Does Label Smoothing Help?"
Pre-registered protocol for an independent PyTorch replication of Müller, Kornblith & Hinton (NeurIPS 2019; arXiv:1906.02629), covering three claims: the calibration effect of label smoothing (C1), its architecture-dependent accuracy effect (C2), and the distillation penalty for students of label-smoothed teachers (C3).
Authors
- Harsh C Parikh (ORCID: https://orcid.org/0009-0008-9625-9637)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24235920
- Primary Topic
- Machine Learning and Data Classification
- Type
- preprint