Infusing Causal Knowledge Into Deep Neural Networks

Coenraads, Wester (2021) Infusing Causal Knowledge Into Deep Neural Networks. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2021_CoenraadsW.pdf Download (2MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (127kB)

Abstract

Deep neural networks can offer high accuracy on a variety of tasks, but are often found to be uninterpretable, unexplainable and exploitable in part because of their reliance on correlations learned from training data. Models built on causal relations, rather than correlations, are in theory more accurate and explainable. We present several novel methods to infuse causal knowledge into neural networks: the TCAV Loss Function, the Multi-Task Concept Network, the Graph-CNN and the Leaky Semantic Bottleneck Network. Our experimental results show that each of these methods performs better than a baseline model on a dataset watermarked with artificial correlations. Out of our methods, the Leaky Semantic Bottleneck Network (LSBN) attains the best results and is further tested on larger datasets, achieving an accuracy of 72% compared to a baseline of 60%. Based on the performance of each of the tested models with infused causal relations, we conclude that models infused with causal relations are more accurate than models that are built purely on correlations learned from the data.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Jaeger, H. and Gaydadjiev, G.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	30 Jul 2021 05:45
Last Modified:	30 Jul 2021 05:45
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/25514

Actions (login required)

View Item