Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Machine Unlearning under Overparameterization
Authors: Jacob Block, Aryan Mokhtari, Sanjay Shakkottai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide exact and approximate unlearning guarantees and demonstrate that an implementation of our framework outperforms existing baselines across various unlearning experiments. We show strong performance of our method across three experimental settings: Data Poisoning, Multi-Class Label Erasure, and Representation Collapse, using natural and interpretable unlearning metrics to compare our method against existing baselines. |
| Researcher Affiliation | Collaboration | Jacob L. Block UT Austin EMAIL Aryan Mokhtari UT Austin & Google Research EMAIL Sanjay Shakkottai UT Austin EMAIL |
| Pseudocode | Yes | Algorithm 1 Min Norm-OG |
| Open Source Code | Yes | Code is publicly released and the datasets we use (Tiny Image Net, CIFAR-10) are publicly available. |
| Open Datasets | Yes | We use the CIFAR-10 [30] and Tiny Image Net [31] datasets, creating red, green, and gray copies of each image. We use the CIFAR-10 [30] and Tiny Image Net [31] datasets in our experiments. |
| Dataset Splits | Yes | We train a shallow neural network on retain samples (xr, yr) Dr with yr = sin(xr) and forget samples (xf, yf) Df with yf = 1.5, over input domain X = [ 5π, 5π] R. The retain set Dr contains all image content classes only in gray, while the forget set Df contains all colors. The retain set Dr consists of images colored according to their assigned class color, while the forget set Df contains randomly colored images. For evaluation, we label heldout test images by color and assess unlearning via color-label accuracy, testing if the unlearning methods can collapse the original model into just a color classifier. |
| Hardware Specification | Yes | All experiments were run on either a single NVIDIA A40 GPU or a single NVIDIA H200 GPU. All training and parameter searches were performed on a cluster of NVIDIA GH200 GPUs. All training was performed on a cluster of NVIDIA H200 GPUs. |
| Software Dependencies | No | For each method, we use the Adam W optimizer with learning rate η on different effective loss functions. For CIFAR-10, we train for 100 epochs using the SGD optimizer with an initial learning rate of 3 10 2, weight decay of 5 10 4, momentum of 0.9, and batch size of 256. For Tiny Image Net, we initialize the Res Net-50 architecture with Image Net-pretrained weights [33] from the torchvision library. |
| Experiment Setup | Yes | For each seed, we randomly sample 50 retain set points (xr, yr) Dr with yr = sin(xr) and 5 forget set points (xf, yf) Df with yf = 1.5, over the input domain X = [ 5π, 5π] R. We initially train the poisoned model on all the samples using the Adam W optimizer with a learning rate of 10 3 over 100,000 epochs. For CIFAR-10, we train for 100 epochs using the SGD optimizer with an initial learning rate of 3 10 2, weight decay of 5 10 4, momentum of 0.9, and batch size of 256. The learning rate is reduced to 3 10 3 at epoch 50. Table 6: Hyperparameter settings for each entry in Table 5. |