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

We typically compute aggregate statistics on held-out test data to assess thegeneralization of machine learning models. However, statistics on test dataoften overstate model generalization, and thus, the performance of deployedmachine learning models can be variable and untrustworthy. Motivated by theseconcerns, we develop methods to automatically discover and correct model errorsbeyond those available in the data. We propose Defuse, a method that generatesnovel model misclassifications, categorizes these errors into high-level modelbugs, and efficiently labels and fine-tunes on the errors to correct them. Togenerate misclassified data, we propose an algorithm inspired by adversarialmachine learning techniques that uses a generative model to find naturallyoccurring instances misclassified by a model. Further, we observe that thegenerative models have regions in their latent space with higher concentrationsof misclassifications. We call these regions misclassification regions and findthey have several useful properties. Each region contains a specific type ofmodel bug; for instance, a misclassification region for an MNIST classifiercontains a style of skinny 6 that the model mistakes as a 1. We can also assigna single label to each region, facilitating low-cost labeling. We propose amethod to learn the misclassification regions and use this insight to bothcategorize errors and correct them. In practice, Defuse finds and correctsnovel errors in classifiers. For example, Defuse shows that a high-performancetraffic sign classifier mistakes certain 50km/h signs as 80km/h. Defusecorrects the error after fine-tuning while maintaining generalization on thetest set.

Defuse: Harnessing Unrestricted Adversarial Examples for Debugging Models Beyond Test Accuracy