[RuleML Webinar] On 31/1: Generating Concept-based and Relational Explanations for Image Classification
2 views
Skip to first unread message
Tomas Kliegr
Jan 30, 2024, 9:31:38 AM1/30/24
to Inductive Logic Programming News
Dear all,
I am forwarding an announcement of a talk on combining relational learning with image classification. The talk will be held as part of the RuleML webinar series. I hope this will be of interest to some on this mailing list.
----------------------------------------------------------------------------------------------------------------------------------------
TITLE: Generating Concept-based and Relational Explanations for Image Classification
DATE AND TIME: January 31 starting at 3 pm CET (30 minutes incl. discussion)
SPEAKER: Bettina Finzel (Cognitive Systems group, University of Bamberg, Germany).
CONNECTION INFORMATION: ZOOM
https://cesnet.zoom.us/j/99923246896?pwd=NDl2RWI0QmV5MjF0NG9mYkpSUU53QT09
Meeting ID: 999 2324 6896
Passcode: 671055
TALK ABSTRACT: The prevalent way to explain decisions of image classifiers such as Convolutional Neural Networks (CNNs) are saliency maps, better known as heat maps. Visualizations of this form are, however, limited in their expressiveness with respect to making explicit what strategies a CNN has used to classify sample images. In this talk, I will introduce CoReX, a framework that fuses saliency-based explanations with rule-based post-hoc explanations. I show that this framework builds on top of the state-of-the-art in extracting comprehensible concepts from neural network models and that CoReX extends these visual explanations by computing logic-based, relational information and providing the means for more in-depth and interactive exploration and evaluation of neural network decisions.
SPEAKER ACADEMIC BIO: Bettina Finzel worked from 2019 to 2023 as a research associate on methods of explainable artificial intelligence for medicine at the Cognitive Systems group at the University of Bamberg. Since 2023, she has worked as a teaching assistant at the Cognitive Systems group and is currently finalizing her doctoral thesis on explainable machine learning under the supervision of Prof. Ute Schmid.
----------------------------------------------------------------------------------------------------------------------------------------
----------------------------------------------------------------------------------------------------------------------------------------
TITLE: Generating Concept-based and Relational Explanations for Image Classification
DATE AND TIME: January 31 starting at 3 pm CET (30 minutes incl. discussion)
SPEAKER: Bettina Finzel (Cognitive Systems group, University of Bamberg, Germany).
CONNECTION INFORMATION: ZOOM
https://cesnet.zoom.us/j/99923246896?pwd=NDl2RWI0QmV5MjF0NG9mYkpSUU53QT09
Meeting ID: 999 2324 6896
Passcode: 671055
TALK ABSTRACT: The prevalent way to explain decisions of image classifiers such as Convolutional Neural Networks (CNNs) are saliency maps, better known as heat maps. Visualizations of this form are, however, limited in their expressiveness with respect to making explicit what strategies a CNN has used to classify sample images. In this talk, I will introduce CoReX, a framework that fuses saliency-based explanations with rule-based post-hoc explanations. I show that this framework builds on top of the state-of-the-art in extracting comprehensible concepts from neural network models and that CoReX extends these visual explanations by computing logic-based, relational information and providing the means for more in-depth and interactive exploration and evaluation of neural network decisions.
SPEAKER ACADEMIC BIO: Bettina Finzel worked from 2019 to 2023 as a research associate on methods of explainable artificial intelligence for medicine at the Cognitive Systems group at the University of Bamberg. Since 2023, she has worked as a teaching assistant at the Cognitive Systems group and is currently finalizing her doctoral thesis on explainable machine learning under the supervision of Prof. Ute Schmid.
----------------------------------------------------------------------------------------------------------------------------------------
ABOUT RULEML WEBINAR SERIES
RuleML webinars are a spin off of the RuleML conference series. The webinar was started in 2015. To date, it hosted over fifty talks covering a variety of topics related to research and applications of rule technologies.
RuleML is now part of RuleML+RR, held under the DeclarativeAI conference series.
* Past RuleML talks: https://github.com/RuleML/ruleml-website/blob/master/talks/README.md
* Slides from the last RuleML webinar: Rule learning for recommender systems by Len Feremans (Adrem Data lab at the University of Antwerp): https://github.com/RuleML/ruleml-website/blob/master/talks/LensFeremans-RuleLearningForRecommendation-2023-10-25.pdf
----------------------
----------------------
Tomas Kliegr
Associate professor
Department of Information and Knowledge Engineering
VSE Prague
Czech Republic
Reply all
Reply to author
Forward
0 new messages