Thread by @hima_lakkaraju

Regulating #AI is important, but it can also be quite challenging in practice. Our #ICML2023 paper highlights the tensions between Right to Explanation & Right to be Forgotten, and proposes the first algorithmic framework to address these tensions arxiv.org/pdf/2302.04288… [1/N]

@SatyaIsIntoLLMs @Jiaqi_Ma_ Multiple regulatory frameworks (e.g., GDPR, CCPA) were introduced in recent years to regulate AI. Several of these frameworks emphasized the importance of enforcing two key principles ("Right to Explanation" and "Right to be Forgotten") in order to effectively regulate AI [2/N]

While Right to Explanation ensures that individuals who are adversely impacted by algorithmic outcomes are provided with an actionable explanation, Right to be Forgotten allows individuals to request erasure of their personal data from databases/models of an organization [3/N]

While several regulatory frameworks emphasize the importance of enforcing both the aforementioned principles, it is unclear if there are any trade-offs between these principles and/or if it is even feasible to simultaneously enforce them in practice? [4/N]

In our #ICML2023 paper, we investigate the tensions that arise when we try to simultaneously enforce both Right to Explanation & Right to be Forgotten in practice. These tensions stem from the characteristics of SOTA ML methods designed to operationalize the two principles. [5/N]

Intuitively, the tension between the two principles stems due to the following: enforcing right to be forgotten may trigger model updates which in turn invalidate previously provided actionable explanations that end users may act upon, thus violating right to explanation. [6/N]

We find that actionable (counterfactual) explanations generated by SOTA ML algorithms become invalid (i.e., acting upon them will no longer result in desired model prediction) once the underlying model is updated in order to accommodate data deletion requests. [7/N]

To address the aforementioned challenges, we propose the first algorithmic framework, RObust Counterfactual Explanations under Right to be Forgotten (ROCERF), to address the tension between the two key regulatory principles -- Right to Explanation and Right to be Forgotten. [8/N]

In particular, we formulate a novel optimization problem to generate actionable (counterfactual) explanations that are robust to model updates triggered due to data deletion requests. [9/N]

We also derive an efficient algorithm to handle the combinatorial complexity of the above optimization problem. To this end, we build on ideas and techniques from existing literature on actionable (counterfactual) explanations, unlearning, as well as leave-k-out estimation [10/N]

We also derive theoretical results which demonstrate that the actionable explanations output by our method are provably robust to model updates triggered due to worst-case data deletion requests with bounded costs for linear models and certain classes of non-linear models. [11/N]

Our experimental results with multiple real-world datasets also validate the effectiveness of our method in comparison with other SOTA actionable explanation methods as well as their general-purpose robust counterparts. [12/N]

Last but not the least, this work was the result of a ton of hard work by two amazing students from our @ai4life_harvard group -- @SatyaIsIntoLLMs and @Jiaqi_Ma_ [13/N]

Please do reach out if you have any feedback/inputs on this work or thoughts about future research directions. Thank you for your time and attention! [N/N]

https://twitter.com/hima_lakkaraju/status/1653468251626565633