Sparse Machine Unlearning via Parameter Localization

This direction is merged into another direction: catastrophic forgetting in unlearning. Its last version stays readable and citable here.

Version 1 of 135 papers41 conceptsfrozen 2026-10-04

Machine unlearning—the process of scrubbing specific data, concepts, or behaviors from a pre-trained model—traditionally presents a binary choice: the "gold standard" of expensive retraining from scratch, or fast, approximate fine-tuning that often "lobotomizes" the model, destroying its general utility.

Recent research has converged on a third way: sparsity. By identifying and modifying only the tiny fraction of parameters responsible for specific memories, researchers can achieve near-exact unlearning without the collateral damage of global weight updates.

The Intuition of Localization

The core insight driving this field is that knowledge in large-scale models is not uniformly distributed. Instead, specific facts, styles, and data signatures are often localized within sparse sub-networks or "circuits."

Harry J. Davies and colleagues demonstrated in Targeted Angular Reversal of Weights (TARS) that factual knowledge is stored with extreme sparsity. Similarly, Kaiyuan Deng showed in Forget-It-All that unwanted concepts in diffusion models are localized within less than 0.3% of total neurons.

When we update every parameter in a model to forget one piece of information, we introduce noise into millions of unrelated circuits. By restricting updates to the "culprit" weights, we preserve the model's general intelligence.

Identifying the Culprit Weights

To unlearn surgically, one must first find the right parameters to edit. Several mechanisms have emerged to map specific data to specific weights:

  • Gradient Localization: Niklas Stoehr and Mitchell Gordon found that updating less than 1% of weights with the highest gradients for memorized content can effectively neutralize specific training data Localizing Paragraph Memorization in Language Models.
  • Generalized Influence Functions (GIFs): Hyeonsu Lyu and others introduced GIFs, which focus updates on the most relevant 5% of parameters to avoid the accumulation of approximation errors common in standard influence functions.
  • Singular Value Decomposition (SVD): The SEMU framework by Marcin Sendera uses SVD to isolate singular vectors that contribute most to a specific objective, modifying as little as 0.5% of the model.
  • Weight Attribution: Jinghan Jia developed WAGLE, which uses sparse parameter masking to isolate influential weights, preventing the collapse of general reasoning capabilities during the unlearning process.

Sparsity as a Buffer Against Utility Collapse

A recurring problem in sequential unlearning—where a model must forget multiple concepts over time—is "utility collapse." Justin Lee and Zheda Mai observed that sequential updates cause cumulative parameter drift Continual Unlearning for Text-to-Image Diffusion Models.

Sparsity acts as a protective buffer. By "pruning first, then unlearning," as suggested by Jinghan Jia in Model Sparsification Can Simplify Machine Unlearning, the gap between approximate unlearning and full retraining significantly closes. In modular architectures like Mixture-of-Experts (MoE), this is even more pronounced: Haomin Zhuang showed that updating just one expert per layer (0.06% of parameters) is more effective than broader updates SEUF.

Theoretical Foundations

This empirical success is supported by new theoretical frameworks. Sadhika Malladi and Stanley Wei proved that unlearning is actually easier when a model is fine-tuned for a sparse downstream task Provable unlearning in topic modeling and downstream tasks. Because these tasks only rely on a subset of features, the model can tolerate the removal of large volumes of training data related to irrelevant features without losing performance.

Furthermore, Linda Lu and Ayush Sekhari argue for System-Aware Unlearning, which reframes unlearning from a subtraction problem into a sparse selection problem. If a system only trains on a core subset of data, deleting points outside that set requires zero computation, drastically reducing the overhead of large-scale data removal.

Summary

The shift toward sparse, localized updates suggests that the "memory" of a neural network is not a monolithic block but a collection of overlapping, sparse sub-structures. By treating unlearning as a surgical intervention rather than a blunt fine-tuning process, we can satisfy the "right to be forgotten" while maintaining the utility of the world's most powerful AI models.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sparse_machine_unlearning_via_parameter__4aa578c2,
  title = {Sparse Machine Unlearning via Parameter Localization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparse-machine-unlearning-via-parameter-localization/rdr_4aa578c2ecb647b98d38a55a4fc14047},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_sparse_machine_unlearning_via_parameter__4aa578c2_v1,
  title = {Sparse Machine Unlearning via Parameter Localization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparse-machine-unlearning-via-parameter-localization/rdr_4aa578c2ecb647b98d38a55a4fc14047/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}