DISCERN authors: Distributional Synthetic Replay for Retain-Free Incremental Machine Unlearning

DISCERN authors: Distributional Synthetic Replay for Retain-Free Incremental Machine Unlearning

The authors present DISCERN, a class unlearning method designed for deletion requests that arrive one after another. In incremental class unlearning, a model must forget each removed class, keep previously forgotten classes forgotten, and preserve the classes that remain. DISCERN replaces access to the retain set—the training data for classes being kept—with synthetic embeddings sampled from per-class parametric distributions. Rather than assume every class follows the same kind of distribution around its centroid, the method tests several Euclidean and directional families and selects among them. The authors say this choice reflects differences in latent geometry between transformer and convolutional networks. In experiments on MNIST, CIFAR-10, and CIFAR-100, across four backbones, the authors report that DISCERN was the most stable method in incremental class unlearning on transformers. On ViT-B/16, after 20 sequential deletions, its harmonic mean was 0.797; the strongest method in the simultaneous setting fell from 0.918 to 0.048. A paired ablation, the authors say, found that per-class distribution selection reduced resurfacing of previously forgotten classes. They characterize this as stabilizing accumulated forgetting, not improving every result uniformly. The reported findings concern the datasets, backbones, and unlearning settings evaluated.

Publication: 31ba521c-b99b-4f9e-8bfa-3bc55c93c428

attribution — pass: The reading attributes DISCERN and all reported results to the paper's authors, matching the arXiv listing. Results are framed as what the authors report or say, not as established fact.

Authors:Nicolás Segura-Kunsagi, Juan Luis Suárez-Díaz, Iván Sevillano-García, Francisco Herrera

summary — uncertain: The summary is faithful, but a non-specialist cannot readily follow the central idea or result. It never says plainly what machine unlearning is: removing a training data's influence without retraining from scratch. Key terms go unexplained in plain language, including synthetic embeddings, parametric distributions, Euclidean/directional families, latent geometry, backbones and ablation. Above all, it does not say what the harmonic mean measures, so the headline figures 0.797 and 0.048 are hard to interpret. Smallest repair: add a one-clause definition of unlearning and of what the harmonic mean balances, or ensure term-library tooltips cover these.

Machine unlearning aims to remove the influence of specific training data from a trained model without retraining from scratch.

date — pass: The originalPublished value of 2026-11-05 matches the v1 submission date on the arXiv page. The later feed timestamp is correctly not used.

[Submitted on 5 Nov 2026]

support — pass: Each claim matches the abstract: retain-set replacement, stability in ICU on transformers, the ViT-B/16 figures, and the ablation characterization. The limitations scope results to the tested datasets and backbones without overreach. The question is a genuine open issue and not a leading claim. No unsupported benefits or costs are implied.

DISCERN is competitive when classes are removed at once and is the most stable method in ICU on transformers

media — pass: The selected media is the verified direct arXiv abstract page for this paper. That is the most direct access to the source.