Auditing Deep Neural Networks and Other Black-box Models
Date
2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Producer
Director
Performer
Choreographer
Costume Designer
Music
Videographer
Lighting Designer
Set Designer
Crew Member
Funder
Rehearsal Director
Concert Coordinator
Advisor
Moderator
Panelist
Alternative Title
Department
Haverford College. Department of Computer Science
Type
Thesis
Original Format
Running Time
File Format
Place of Publication
Date Span
Copyright Date
Award
Language
eng
Note
Table of Contents
Terms of Use
Rights Holder
Access Restrictions
Open Access
Terms of Use
Tripod URL
Identifier
Abstract
In this era of self-driving cars, smart watches, and voice-commanded speakers, machine learning is ubiquitous. Recently, deep learning has shown impressive success in solving many machine learning problems related to image data and sequential data - with the result that people are frequently impacted by deep learning models on a daily basis. However, how do we judge if these models are fair, and how do we discover what information is important when making a decision? And as APIs become ever-more common, how do we determine this information if we do not have access to the model itself? We developed a novel technique called Gradient Feature Auditing which gradually obscures information from a data-set and evaluates how a model's predictions change as yet more of that information is obscured. This allows a deeper investigation of what information and features are actually used by machine learning models when making predictions. Throughout our experiments, we apply Gradient Feature Auditing on multiple data-sets using several popular modeling techniques (linear SVMs, C4.5 decision trees, and shallow feed-forward neural networks) to provide evidence that Gradient Feature Auditing indeed affords deeper insight into what information a model is using.