Title:	ODIN: A FRAMEWORK FOR ERROR DIAGNOSIS OF DNNS
Author(s):	Rocio Nahime Torres, Federico Milani, Niccolò Zangrando and Piero Fraternali
ISBN:	978-989-8704-42-9
Editors:	Yingcai Xiao, Ajith Abraham, Guo Chao Peng and Jörg Roth
Year:	2022
Edition:	Single
Keywords:	Error Diagnosis, Computer Vision, Model Evaluation, Black-Box, Error Categorization
Type:	Poster
First Page:	241
Last Page:	243
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Methods for the analysis of Deep Neural Networks (DNNs) use two approaches: open- and black-box. The formers apply explainability techniques to assess the relationship between the input, the inner layers, and the output. Black-box approaches exploit semantic properties of the input additional w.r.t. those used for training. They enable the characterization of performance metrics and errors and help scientists identify the input features responsible for prediction failures and focus their model improvement efforts. In this demo we provide a guided tour of the functionalities of ODIN, an open-source framework for black-box DNN model diagnosis. We show how to investigate the behavior of classification and object detection models using a broad range of metrics and performance reports, easily extensible by the model developer.
	Go Back