English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Influence functions approximate the "influences" of training data-points fortest predictions and have a wide variety of applications. Despite thepopularity, their computational cost does not scale well with model andtraining data size. We present FastIF, a set of simple modifications toinfluence functions that significantly improves their run-time. We usek-Nearest Neighbors (kNN) to narrow the search space down to a subset of goodcandidate data points, identify the configurations that best balance thespeed-quality trade-off in estimating the inverse Hessian-vector product, andintroduce a fast parallel variant. Our proposed method achieves about 80Xspeedup while being highly correlated with the original influence values. Withthe availability of the fast influence functions, we demonstrate theirusefulness in four applications. First, we examine whether influentialdata-points can "explain" test time behavior using the framework ofsimulatability. Second, we visualize the influence interactions betweentraining and test data-points. Third, we show that we can correct model errorsby additional fine-tuning on certain influential data-points, improving theaccuracy of a trained MultiNLI model by 2.5% on the HANS dataset. Finally, weexperiment with a similar setup but fine-tuning on datapoints not seen duringtraining, improving the model accuracy by 2.8% and 1.7% on HANS and ANLIdatasets respectively. Overall, our fast influence functions can be efficientlyapplied to large models and datasets, and our experiments demonstrate thepotential of influence functions in model interpretation and correcting modelerrors. Code is available athttps://github.com/salesforce/fast-influence-functions

FastIF: Scalable Influence Functions for Efficient Model Interpretation and Debugging