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..
CoUn: Empowering Machine Unlearning via Contrastive Learning
Authors: Yasser Khalil, Mehdi Setayesh, Hongliang Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across various datasets and model architectures show that Co Un consistently outperforms state-of-the-art MU baselines in unlearning effectiveness. Additionally, integrating our CL module into existing baselines empowers their unlearning effectiveness. |
| Researcher Affiliation | Industry | Huawei Noah s Ark Lab, Montreal, Canada EMAIL |
| Pseudocode | Yes | Appendix B: Pseudo-Code and Py Torch implementation of Co Un. B.1 Pseudo-Code Algorithm 1 details our proposed unlearning method, which leverages CL and supervised learning. |
| Open Source Code | Yes | We uploaded Co Un code in the supplementary material and will be made publicly available. |
| Open Datasets | Yes | We evaluate Co Un on three datasets: CIFAR-10/100 [36] and Tiny Image Net [37], using three model architectures: Res Net-18 [38], VGG-16 [39], and Vi T [40]. |
| Dataset Splits | Yes | Given a dataset D, partitioned into forget data Du and retain data Dr = D \ Du, the goal of MU is to transform an Original model θo, trained on D, into an unlearned model θu that effectively removes the influence of Du. We define the forget data ratio as: |Du|/|D| * 100. |
| Hardware Specification | Yes | All experiments are implemented using the Py Torch platform [43] and run using NVIDIA Tesla V100 GPUs. |
| Software Dependencies | No | All experiments are implemented using the Py Torch platform [43] and run using NVIDIA Tesla V100 GPUs. The paper mentions 'Py Torch platform' but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | For training the Original and Retrain models: we follow prior work [8, 9, 10] by using an initial learning rate of 0.1, which is reduced by a factor of 10 at 50% and 75% of the total 182 training epochs. The batch size is set to 256. An SGD optimizer is used with an initial learning rate of 0.1 and a multi-step learning rate scheduler that reduces the learning rate by a factor of 10 at 50% and 75% of the training epochs. Momentum is set to 0.9, and weight decay is set to 5e-4. For unlearning: all methods are run for 50 epochs. We used the SGD optimizer with a learning rate tuned within the range of [0.01, 0.1] for each MU method. A cosine annealing learning rate scheduler is used with a minimum learning rate set to 1e-4. Momentum is set to 0.9, and weight decay is set to 5e-4. |