Research map / Optimization / Implicit bias and generalization
Loss is Not Behavior: A Unified Output-Space Analysis of Gradient-Based Machine Unlearning
Xingjian Zhao, Mohammad Mohammadi Amiri, Malik Magdon-Ismail
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Analyzes gradient-based machine unlearning in output space, arguing that matching forget/retain losses does not imply retraining-like behavior.
Problem
Gradient-based unlearning methods optimize forget and retain losses, but similar losses do not guarantee output-space behavior matching a model retrained on only the retain data.
Approach
A unified output-space analysis of how a gradient-based unlearner deviates from the retrain model, rather than judging unlearning by loss values alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use matched forget loss or weights as the evaluation criterion for unlearners | gradient-based unlearners · MNIST class forgetting | output RMS gap on forget set: SimNPO 0.009 vs ProjGD 0.509 at similar forget CE (4.368-4.404) | Hurt |
| switch from random to class forgetting | GradDiff, NPO, SimNPO · MNIST | RMS_f gap from retrained model: residual norm 5.28 -> 24.25; RMS_f scaled by 1.8-3.8x | Hurt |
| use retained-Hessian leverage geometry instead of Euclidean Gram for gap prediction (ablation) | GradDiff gap predictor · MNIST random forget set, retain anchors | Pearson correlation: 0.40 -> 0.72 | Improved |
| use retained-Hessian leverage geometry instead of Euclidean Gram for gap prediction (ablation) | SimNPO gap predictor · MNIST class forget set, forget set | Pearson correlation: -0.03 -> 0.98 | Improved |
| use projected-gradient dynamics (ProjGD) with unprojected predictor | ProjGD · MNIST and CIFAR-10 head | actual output gap vs prediction: direction correlation 0.79-0.97, actual gap about 7-15x smaller | Mixed |
| remove retain anchoring (Pure GA, IMU) instead of GradDiff-style retain term (ablation) | Pure GA and IMU · MNIST and CIFAR-10 | output KL from retrained model: 0.7-3 nats vs at most 10^-5 | Hurt |
| use SimNPO default beta 2.5 instead of NPO beta 0.1 in class forgetting (ablation) | SimNPO vs NPO · MNIST class forgetting | RMS_f gap: 0.032 -> 0.009 (saturated weights switch off forget term) | Mixed |
| set per-set rho from leverage rule rho = s n / lambda_max(K_ff) with s=0.3 | GradDiff/NPO/SimNPO · 330 MNIST forget sets | optimization stability | Improved |
| use leverage K-channel scores computed at the original model as example-level diagnostic | TOFU Llama-3.2-1B-Instruct LoRA chart · TOFU | Spearman rank correlation with disturbed retain examples: 0.26-0.49 (retain), 0.14-0.29 (forget) | Improved |
| use LoRA-restricted curvature leverage scores to predict gap size | TOFU diagnostics · TOFU | gap size prediction | No clear effect |
| initialize at the original model rather than the retrained reference for the local predictor (ablation) | local gap predictor · MNIST | prediction validity | Hurt |
Surprising findings
- Similar forget and retain losses do not necessarily imply similar output-space behavior to the retrained model.
Closest papers by meaning
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.91)
- The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation (NeurIPS 2026 · similarity 0.90)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.90)
- Ascent Fails to Forget (NeurIPS 2025 · similarity 0.90)
- Efficient Utility-Preserving Machine Unlearning with Implicit Gradient Surgery (NeurIPS 2025 · similarity 0.89)
- Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
