FastIF: Scalable Influence Functions for Efficient Model Interpretation and Debugging

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10333–10350 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10333 FASTIF: Scalable Influence Functions for Efficient Model Interpretation and Debugging Han Guo1,2,3 Nazneen Fatema Rajani1 Peter Hase3 Mohit Bansal3 Caiming Xiong1 1Salesforce Research 2Carnegie Mellon University 3UNC Chapel Hill {nazneen.rajani, cxiong}@salesforce.com hanguo@cs.cmu.edu, {peter, mbansal}@cs.unc.edu Abstract Influence functions approximate the “influ- ences” of training data-points for test pre- dictions and have a wide variety of applica- tions. Despite the popularity, their computa- tional cost does not scale well with model and training data size. We present FASTIF, a set of simple modifications to influence functions that significantly improves their run-time. We use k-Nearest Neighbors (kNN) to narrow the search space down to a subset of good can- didate data points, identify the configurations that best balance the speed-quality trade-off in estimating the inverse Hessian-vector prod- uct, and introduce a fast parallel variant. Our proposed method achieves about 80X speedup while being highly correlated with the original influence values. With the availability of the fast influence functions, we demonstrate their usefulness in four applications. First, we ex- amine whether influential data-points can “ex- plain” test time behavior using the framework of simulatability. Second, we visualize the in- fluence interactions between training and test data-points. Third, we show that we can cor- rect model errors by additional fine-tuning on certain influential data-points, improving the accuracy of a trained MultiNLI model by 2.5% on the HANS dataset. Finally, we experiment with a similar setup but fine-tuning on data- points not seen during training, improving the model accuracy by 2.8% and 1.7% on HANS and ANLI datasets respectively. Overall, our fast influence functions can be efficiently ap- plied to large models and datasets, and our ex- periments demonstrate the potential of influ- ence functions in model interpretation and cor- recting model errors.1 1 Introduction Language understanding systems are becoming ever more powerful with the recent advances in 1Code is available at https://github.com/ salesforce/fast-influence-functions. S-test KNN Training Dataset Subset of the Dataset Influences Computing Influences Most Influential Fast Estimation Model Test Data Model Test Data Model Test Data Figure 1: Workflow of FASTIF w.r.t. a test data-point. First a subset of data-points are selected from the en- tire training set using kNN to reduce search space, then the inverse Hessian-vector product (stest) is estimated based on Sec. 3.3. The influence values of data-points are computed using the outputs from these two steps. Finally, the most influential data-point(s) are returned. large-scale pre-training. As these systems become widely adopted for real-world applications, the abil- ity to interpret the model decisions grows increas- ingly important. An array of interpretation methods now exist for shedding light on models’ decision- making processes, with the bulk of work focusing on estimating feature importance (Li et al., 2016a; Ribeiro et al., 2016b; Lundberg and Lee, 2017; Sun- dararajan et al., 2017a; Ribeiro et al., 2018a) and interpreting internal model representations (Don- nelly and Roegiest, 2019; Mu and Andreas, 2020; De …

Content truncated to 500 words