How to Know When Machine Learning Does Not Now
It is becoming increasingly important to understand how a prediction made by a Machine Learning model is informed by its training data.
adversarial-learning interpretability uncertainty adversarial-examples active-learning outlier-detection anomaly-detection semi-supervised-learning tutorial article research paper arxiv:1803.04765

This post will outline an approach we call Deep k-Nearest Neighbors [Papernot and McDaniel] that attempts to address this issue. We will also explore a loss function we recently introduced to shine some light on how models structure their internal representations [Frosst et al.]. We’ll illustrate how the Deep k-Nearest Neighbors (DkNN) helps recognize data that is not from the training distribution by observing anomalies in the internal representation of such data.

Don't forget to tag @npapernot , @nickfrosst in your comment, otherwise they may not be notified.

Authors
Research Engineer @ Google -- Singer @ Good Kid
Share this project
Similar projects
AllenNLP Interpret
A Framework for Explaining Predictions of NLP Models
Adversarial Training Improves Product Discovery
Method automatically generates meaningful negative training examples for deep-learning model.
Transfer Learning using Adversarially Robust ImageNet Models
We explore adversarial robustness as an avenue for training computer vision models with more transferrable features.
An Introduction to Adversarial Examples in Deep Learning
This report provides an intuitive introduction to adversarial examples, discusses a wide variety of different adversarial attacks and, most notably, ...