Topics — What a work is about. Nucleation holds whatever reading the owner files here, whether or not an Anthology of the SOTA topic could also hold it; instruction for ML practice belongs to the anthology (ADR-048). The words are seeded from the clusters the reading-time triage of 2026-09-25 found, and grow by decision.
Model comparison — choosing between models by their evidence — the marginal likelihood and Bayes factors, the Savage–Dickey density ratio for nested models, Occam factors and the evidence framework, Bayesian model reduction and structure learning, and information criteria and free energies for singular models (ADR-026).
3 of 191 CASE documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| CASE-041 v2 | FID and Projected GANs: one ImageNet-trained probe family used both to train a generator and to evaluate it | The owner pointed to FID at U50 and proposed FID with Projected GANs as a paired case study at U51. A184 framed the pair: FID evaluates a distribution through probes, and Projected GANs learn one under pressure from probes. The anthology’s readings supply the fact the exchange lacked. The Projected GAN discriminator is ImageNet-trained, FID’s features are ImageNet class evidence, and Kynkäänniemi et al. show that such a generator can match FID while doing worse in an independent feature space and with human judges. | Active |
| CASE-070 | Voita, Sennrich and Titov’s BLEU-tied translation systems that differ on keeping ты or вы across sentences | Two English-to-Russian subtitle translation systems score the same BLEU (32.40 and 32.38, not significantly different). On a contrastive test built from held-out films, which asks whether a system keeps the T or V form of “you” consistent across sentences, one is at chance (50.0%) and the other is right in 81.6% of cases. A documented instance, on a pragmatic distinction, of two realizations equal under a standard probe family and different under an independently built one. An ML paper that neither record holds. | Active |
| CASE-071 v2 | Kovacs, Deutsch and Freitag’s metric bias in MBR decoding: translations a metric selects, that the metric and its neighbours rate better and human raters do not | Minimum Bayes risk and quality-estimation decoding pick, from 128 sampled translations, the one a neural metric scores best. That metric and other neural metrics rate the picks significantly better than greedy decoding; human MQM raters find them no better, and for four system and language-pair combinations significantly worse. CASE-041’s result in text, by selection rather than training. The distinction lost is accuracy, not a pragmatic one; a stipulated English-to-Russian ты/вы variant states the design that would make it pragmatic. | Active |