BROWSE

ITEM VIEW & DOWNLOAD

Monte-Carlo Dropout based Uncertainty Analysis in Input Attributions of Multivariate Temporal Neural Networks

Cited 0 times inthomson ciCited 0 times inthomson ci
Title
Monte-Carlo Dropout based Uncertainty Analysis in Input Attributions of Multivariate Temporal Neural Networks
Other Titles
다변량 시간 신경망의 입력 기여도에 대한 몬테카를로 드랍아웃 기반의 불확실성 분석
Author
Lee, Ginkyeng
Advisor
Kim, Kwang In
Keywords
ExplainableAI; uncertainty
Issue Date
2020-02
Publisher
Graduate School of UNIST
Abstract
As deep learning has grown fast, so did the desire to interpret deep learning black boxes. As a result, many analysis tools have emerged to interpret it. Interpretation in deep learning has in fact popularized the use of deep learning in many areas including research, manufacturing, finance, and healthcare which needs relatively accurate and reliable decision making process. However, there is something we should not overlook. It is uncertainty. Uncertainties of models are directly reflected in the results of interpretations of model decision as explaining tools are dependent to models. Therefore, uncertainties of interpreting output from deep learning models should be also taken into account as quality and cost are directly impacted by measurement uncertainty. This attempt has not been made yet. Therefore, we suggest Bayesian input attribution rather than discrete input attribution by approximating Bayesian inference in deep Gaussian process through dropout to input attribution in this paper. Then we extract candidates that can sufficiently affect the output of the model, taking into account both input attribution itself and uncertainty of it.
Description
Department of Computer Science and Engineering
URI
Go to Link
Appears in Collections:
CSE_Theses_Master
Files in This Item:
Monte-Carlo Dropout based Uncertainty Analysis in Input Attributions of Multivariate Temporal Neural Networks.pdf Download

find_unist can give you direct access to the published full text of this article. (UNISTARs only)

Show full item record

qrcode

  • mendeley

    citeulike

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

MENU