Title: SUPREME: An Open-Source Multi-GPU Framework for Reproducible Evaluation of Image Unlearning Methods

Abstract: Machine unlearning aims to eliminate the impact of specific training data on a model without the need for complete retraining. However, assessing the efficacy of these methods is resource-intensive, as it necessitates repeating the training, unlearning, and evaluation processes across various random seeds. Currently, available frameworks for image classification unlearning are restricted to single-GPU execution, which significantly hampers the ability to evaluate multiple seeds within a practical timeframe. To address this limitation, we present SUPREME, an open-source framework designed to distribute these computational stages across multiple GPUs. SUPREME introduces three primary innovations: a registry-based architecture that facilitates the integration of new methods, metrics, models, and scenarios; a multi-GPU infrastructure capable of supporting various accelerators and precision configurations; and a practical demonstration utilizing the Pins Face Recognition dataset with ResNet18 and ViT models, covering both full-class and random-sample unlearning across ten seeds. The framework can be accessed at https://github.com/pedroandreou/supreme-unlearning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC