One of the biggest criticisms of the field of post hoc #XAI is that each method "does its own thing", it is unclear how these methods relate to each other & which methods are effective under what conditions. Our #NeurIPS2022 paper provides (some) answers to these questions. [1/N] 
In our #NeurIPS2022 paper, we unify eight different state-of-the-art local post hoc explanation methods, and show that they are all performing local linear approximations of the underlying models, albeit with different loss functions and notions of local neighborhoods. [2/N]
By doing so, we are able to explain the similarities & differences between these methods. These methods are similar in the sense that they all perform local linear approximations of models, but they differ considerably in "how" they perform these approximations [3/N]
Literature on local post hoc explanation methods is also commonly criticized for the fact that the definition of "locality" is often unclear. Our work provides one of the first formal characterizations of the notion of local neighborhoods employed by various SOTA methods. [4/N]
We also introduce a **no free lunch theorem for explanation methods** to demonstrate that no single explanation method can perform optimally across all (notions of) local neighborhoods. This result further highlights the need for choosing among explanation methods [5/N]
We also provide a guiding principle to choose among explanation methods -- a method is considered effective if it perfectly recovers the underlying model when it can i.e., when the (local) underlying model is a member of the explanation's function class. [6/N]
By leveraging the above principle, we theoretically analyze the conditions under which popular explanation methods are effective, and provide recommendations for choosing among them based on the characteristics/distributions of the input variables. [7/N]
Our framework also enables researchers and practitioners to readily construct **new post hoc explanation methods** based on the requirements of their applications and the characteristics of their input variables. More details about this in the paper. [8/N]
We are very grateful to reviewers at #NeurIPS2022 & #ICML Workshop on Interpretable Machine Learning in Healthcare (IMLH) for their invaluable feedback/suggestions which helped us improve the work. This paper also won the best paper award at IMLH.
https://twitter.com/hima_lakkaraju/status/1550865850885873669[9/N]
The link to our paper on arxiv is at arxiv.org/abs/2206.01254. Another detailed thread about this work is at
https://twitter.com/Suuraj/status/1550629869200953344[10/N]
Last but not the least, this work is a result of the hard work and efforts of some amazing students in our group @ai4life_harvard -- #TessaHan and @Suuraj. Many congratulations #TessaHan and @Suuraj !! :) [N/N]
PS If you find the above work exciting & are interested in joining our efforts on #XAI and @trustworthy_ml, do check out our open postdoc positions --
https://twitter.com/hima_lakkaraju/status/1567289686904942593and other openings for new PhD students and interns. Details at himalakkaraju.github.io/openpositions.…
• • •
Missing some Tweet in this thread? You can try to
force a refresh
