Dataset Diversity: A New Approach to Dependable Machine Learning Software

Dataset Diversity: A New Approach to Dependable Machine Learning Software

Shin Nakajima (National Institute of Informatics, Tokyo)
Date: 
2017/12/15 (Fri) 16:00 to 17:30
Place: 
IS Bldg. Collaboration 7
Group: 
Logic Unit

Title:Dataset Diversity: A New Approach to Dependable Machine Learning Software
Presenter : Shin Nakajima (National Institute of Informatics, Tokyo)

Abstract : Because computation results of machine learning programs
are not known in advance, their software testing often adapts the
metamorphic testing method, which uses pseudo oracles based on data
diversity. The learning problem of neural networks is non-convex
optimization, and thus the approaches with data diversity are
inadequate. We propose a new metamorphic testing method which is based
on dataset diversity and behavioral oracle. We present our experience
to apply the methods to two machine learning tasks, SVM and NN.

Contact
ogata@jaist.ac.jp