Are we making progress in unlearning? Findings from the first NeurIPS unlearning competition

Abstract: We present the findings of the first NeurIPS competition on unlearning, which sought to stimulate the development of novel algorithms and initiate discussions on formal and robust evaluation methodologies. The competition was highly successful: nearly 1,200 teams from across the world participated, and a wealth of novel, imaginative solutions with different characteristics were contributed. In this paper, we analyze top solutions and delve into discussions on benchmarking unlearning, which itself is a research problem. The evaluation methodology we developed for the competition measures forgetting quality according to a formal notion of unlearning, while incorporating model utility for a holistic evaluation. We analyze the effectiveness of different instantiations of this evaluation framework vis-a-vis the associated compute cost, and discuss implications for standardizing evaluation. We find that the ranking of leading methods remains stable under several variations of this framework, pointing to avenues for reducing the cost of evaluation. Overall, our findings indicate progress in unlearning, with top-performing competition entries surpassing existing algorithms under our evaluation framework. We analyze trade-offs made by different algorithms and strengths or weaknesses in terms of generalizability to new datasets, paving the way for advancing both benchmarking and algorithm development in this important area.

• The paper introduces a novel evaluation framework using differential privacy-inspired metrics to measure forgetting quality.
• It benchmarks various methods, demonstrating that alternatives can closely match full statistical rigor with reduced computational cost.
• Top-performing algorithms achieve notable unlearning efficiency with minimal utility loss and strong generalizability across datasets.

Evaluation of Unlearning Algorithms: Insights from the NeurIPS 2023 Competition

The paper "Are we making progress in unlearning? Findings from the first NeurIPS unlearning competition" presents a comprehensive analysis of contributions from the inaugural NeurIPS competition on machine unlearning. This competition aimed to drive innovation in algorithm development for efficient data erasure from machine learning models and to promote robust evaluation methodologies. The competition saw participation from nearly 1,200 teams globally, showcasing a diverse range of solutions. This paper explores the effectiveness of these top solutions, evaluates benchmarking methodologies, and discusses broader implications for the field.

Context and Motivation

The rapid advancement of deep learning has been accompanied by increasing reliance on training data, raising critical issues around legal, privacy, and safety implications due to harmful, incorrect, or outdated data perpetuated in these models. Traditional retraining approaches for data removal are computationally prohibitive, particularly for large models, underscoring the need for efficient unlearning methods. Machine unlearning, defined as efficiently erasing the influence of specific training data subsets, emerges as a crucial research area. However, significant challenges persist in designing such algorithms and developing robust benchmarks to reliably assess their performance.

Competition Overview

The NeurIPS competition focused on a practical unlearning scenario where participants aimed to develop algorithms to remove the influence of a subset of facial images from an age predictor model trained on the CASIA-SURF dataset. The competition adopted a structured evaluation framework that balanced forgetting quality—measured via an empirical estimate of divergence between unlearned and retrained model outputs—with model utility and computational efficiency.

Key Findings and Evaluation Framework

Defining Forgetting Quality

The paper's evaluation framework formalizes forgetting quality using an adaptation of the $(\varepsilon, \delta)$-unlearning notion, inspired by differential privacy. This metric captures the distributional closeness of unlearned and retrained model parameters. Specifically, the framework employs a hypothesis-testing interpretation, estimating $\varepsilon$ based on the failure rates of attacks designed to distinguish unlearned models from retrained models.

Practical Instantiations and Trade-offs

To balance accuracy and computational cost, the framework explores various practical instantiations. The "Full" setup, despite its statistical rigor, is computationally intensive. Hence, the paper investigates alternatives like "Reuse-$N$-$N$" and bootstrapping, revealing that these alternatives can approximate the "Full" setup effectively, thus offering feasible pathways for practical evaluation.

Algorithm Performance and Generalizability

The competition surfaced several novel algorithms that outperform pre-existing state-of-the-art methods in forgetting quality ($\mathcal{F}$-score) and final score after utility adjustment. Top-performing methods, such as those from the teams "Sebastian" and "Fanchuan," demonstrate significant advancements in unlearning with minimal utility loss.

The investigation of generalizability using the FEMNIST dataset indicates that while some competition methods adapt well with minimal tuning, re-tuning is essential for optimal performance across different datasets. Methods like "Sebastian" exhibit strong out-of-the-box performance, underscoring the importance of algorithmic robustness.

Implications and Future Research Directions

The findings highlight meaningful progress in unlearning, evidenced by advancements in algorithmic performance and robustness under the proposed evaluation metrics. However, the study also underscores the need for continuous refinement of evaluation methodologies to ensure they are computationally viable and broadly applicable. Notably, the exploration of generalizability and alternative aggregation strategies for metrics like $\mathcal{F}$ and utility adjustments are pivotal for broader application.

Future research should explore optimizing evaluation frameworks for different unlearning applications, enhancing hyperparameter tuning processes, and developing algorithms that inherently balance efficiency, utility, and privacy. As the landscape of unlearning continues to evolve, fostering collaborations and standardizing benchmarks across diverse application scenarios will be critical for advancing this burgeoning field.

Conclusion

The NeurIPS 2023 competition catalyzed significant strides in unlearning research, with many contributions surpassing existing benchmarks in terms of empirical metrics of forgetting quality and model utility. The paper's proposed evaluation framework and subsequent analyses offer valuable insights and pathways for future advancements. Nevertheless, ongoing efforts are required to refine these methodologies, ensuring that unlearning frameworks remain applicable, scalable, and efficient across various contexts. The community's collective endeavor towards these goals promises to pave the way for more robust, privacy-preserving machine learning models.