Examining the Effects of Deep Learning Model Structure on Model Interpretability for Time-Series Classifications in Fire Research

Author(s)

Abstract

This present work utilizes an interpretability model to understand and explain the decisions of deep learning models. The use of DeepLIFT is proposed and attributions of a study case are obtained. Benchmarking against two other interpretability models, namely Grad-CAM and dCAM, is conducted. Results show that DeepLIFT can provide precise attributions to the model inputs in both temporal and spatial directions. A parametric study is also carried out to understand the effects of deep learning model structure on the attributions obtained from the interpretability model. Ten different convolutional neural network model structures are considered. Three important observations are made: 1) changes in the model structure have minor effects on the attributions in the temporal direction, but 2) they have negligible effects on attributions in the spatial direction, and 3) convolutional layers need to be fixed to avoid attribution discrepancies. It is hoped that this work can contribute to the development of trustworthy deep learning models for the fire research community

Download Paper

Keywords

Trustworthy and responsible AI, Machine learning, First responder preparedness, Fire risk reduction and Fire fighting

Citation

Issues

If you have any questions about this publication or are having problems accessing it, please contact [email protected].