Please use this identifier to cite or link to this item:
Title: Human explainability through an auxiliary Neural Network
Author: Garcia Sánchez, Albert
Director/Tutor: Seguí Mesquida, Santi
Keywords: Xarxes neuronals (Informàtica)
Aprenentatge automàtic
Tesis de màster
Intel·ligència artificial
Neural networks (Computer science)
Machine learning
Masters theses
Artificial intelligence
Issue Date: 30-Jun-2020
Abstract: [en] Explainability in Deep Learning has become a hot topic in recent years due to the necessity of insights and justifications for predictions. Although this field has an extensive range of different approaches, this thesis explores the feasibility of a new methodology that seeks to provide human-interpretable explanations for each sample being processed by a Neural Network. The term black box is often used in the Explainability field, meaning that there is a lack in transparency within the model when processing data. The explored approach tries to deal with the black box by using the outputs of the hidden layers of a Neural Network as inputs for the model responsible for the explanations. This model is another Neural Network that can be seen as an auxiliary Neural Network to the main task. The predicted explanations are formed by a subset of a list of human-designed justifications for the possible outcomes of the main task. Using the predictions from both networks a cross comparison process is also performed in order to build confidence on the main predictions. Results successfully show how a significant proportion of incorrect outputs are questioned thanks to the predicted explanations.
Note: Treballs finals del Màster de Fonaments de Ciència de Dades, Facultat de matemàtiques, Universitat de Barcelona, Any: 2020, Tutor: Santi Seguí Mesquida
Appears in Collections:Màster Oficial - Fonaments de la Ciència de Dades
Programari - Treballs de l'alumnat

Files in This Item:
File Description SizeFormat 
170892.pdfMemòria2.07 MBAdobe PDFView/Open
TFM-UB-master.zipCodi font20.67 MBzipView/Open

This item is licensed under a Creative Commons License Creative Commons