Corrective Machine Unlearning
Authors: Shashwat Goel, Ameya Prabhu, Philip Torr, Ponnurangam Kumaraguru, Amartya Sanyal
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we run experiments with two types of manipulations that can occur in real-world data collection pipelines, and test whether existing methods can unlearn their adverse effects. We first describe these manipulations, and then our dataset, model architecture, and unlearning setup. |
| Researcher Affiliation | Academia | Shashwat Goel EMAIL IIIT Hyderabad ELLIS Institute Tübingen Max Planck Institute for Intelligent Systems Ameya Prabhu University of Oxford Tübingen AI Center Philip Torr University of Oxford Ponnurangam Kumaraguru IIIT Hyderabad Amartya Sanyal University of Copenhagen EMAIL |
| Pseudocode | No | The paper describes the unlearning methods in prose in Section 3.2 and Appendix A.2, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements or links indicating that source code for the described methodology is publicly available. |
| Open Datasets | Yes | We first use the CIFAR10 and CIFAR100 (Krizhevsky et al., 2009) datasets as standard benchmarks in the unlearning literature (Foster et al., 2023; Kurmanji et al., 2023; Chundawat et al., 2023a). We then report poison unlearning results on PCam (Veeling et al., 2018), a binary classification medical imaging dataset, as a potential application. |
| Dataset Splits | Yes | For all results, metrics are computed on the test set containing unseen samples. The mean and standard deviation are reported over 3 seeds. In the interclass confusion evaluation, for CIFAR10, we confuse the Cat and Dog classes, and for CIFAR100, the maple and oak tree classes, to be consistent with the setup of Goel et al. (2023). |
| Hardware Specification | Yes | The setup used for all experiments is a PC with a Intel(R) Xeon(R) E5-2640 2.40 GHz CPU, 128GB RAM and 1 Ge Force RTX 2080 GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch' in reference to a ResNet-9 implementation (Idelbayev, 2018) but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Our standard training procedure A is as follows: We train our models for 4000 steps on CIFAR10, PCAM and 6000 steps on CIFAR100. Each step consists of training on a single batch, and we use a batch size of 512 throughout. We use an SGD optimizer with momentum 0.9 and weight decay 5e-4, a linear scheduler with tmult = 1.25, and warmup steps as 1 100 of the total training steps. The same hyperparameters are used during unlearning unless otherwise specified. |