Towards Unbounded Machine Unlearning
Authors: Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, Eleni Triantafillou
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The above are substantiated through a comprehensive empirical evaluation against previous state-of-the-art. |
| Researcher Affiliation | Collaboration | Meghdad Kurmanji University of Warwick Peter Triantaļ¬llou University of Warwick Jamie Hayes Google Deep Mind Eleni Triantaļ¬llou Google Deep Mind |
| Pseudocode | Yes | We provide pseudocode, training plots and ablations in the Appendix. (Section 3.1) and Algorithm 1 SCRUB in Section 9. |
| Open Source Code | Yes | Our code is available for reproducibility 2. https://github.com/Meghdad92/SCRUB |
| Open Datasets | Yes | We utilize the same two datasets from previous work: CIFAR-10 [Krizhevsky et al., 2009] and Lacuna-10 [Golatkar et al., 2020a], which is derived from VGG-Faces [Cao and Yang, 2015] |
| Dataset Splits | Yes | For the small-scale, we exactly follow the setup in [Golatkar et al., 2020b] that uses only 5 classes from each of CIFAR and Lacuna ( CIFAR-5 / Lacuna-5 ), with 100 train, 25 validation and 100 test examples per class. and In our experiments, the train, test, and validation sizes are 40000, 10000, and 10000 respectively. |
| Hardware Specification | Yes | For scale-up experiments, the code is executed in Python 3.8, on an Ubuntu 20 machine with 40 CPU cores, a Nvidia GTX 2080 GPU and 256GB memory. |
| Software Dependencies | No | For scale-up experiments, the code is executed in Python 3.8, on an Ubuntu 20 machine with 40 CPU cores, a Nvidia GTX 2080 GPU and 256GB memory. (Mentions Python version, but no other key libraries with versions.) |
| Experiment Setup | Yes | For all experiments, we initialize the learning rate at 0.0005 and decay it by 0.1 after a number of min and max steps. [...] We apply a weight decay of 0.1 for small-scale setting and 0.0005 for large scale experiments, with a momentum of 0.9. Finally, we use different batch sizes for the forget-set and the retain-set to control the number of iteration in each direction, i.e the max and the min respectively. We report these in Table 3. |