A Logic-based Approach to Contrastive Explainability for Neurosymbolic Visual Question Answering

Eiter, ThomasInstitute for Logic and Computation, TU Wien, Vienna, Austria (författare)

Visual Question Answering (VQA) is a well-known problem for which deep-learning is key. This poses a challenge for explaining answers to questions, the more if advanced notions like contrastive explanations (CEs) should be provided. The latter explain why an answer has been reached in contrast to a different one and are attractive as they focus on reasons necessary to flip a query answer. We present a CE framework for VQA that uses a neurosymbolic VQA architecture which disentangles perception from reasoning. Once the reasoning part is provided as logical theory, we use answer-set programming, in which CE generation can be framed as an abduction problem. We validate our approach on the CLEVR dataset, which we extend by more sophisticated questions to further demonstrate the robustness of the modular architecture. While we achieve top performance compared to related approaches, we can also produce CEs for explanation, model debugging, and validation tasks, showing the versatility of the declarative approach to reasoning.

Geibinger, TobiasInstitute for Logic and Computation, TU Wien, Vienna, Austria (författare)
Higuera, NelsonInstitute for Logic and Computation, TU Wien, Vienna, Austria (författare)
Oetsch, JohannesInstitute for Logic and Computation, TU Wien, Vienna, Austria(Swepub:hj)oetjoh (författare)
Institute for Logic and Computation, TU Wien, Vienna, Austria (creator_code:org_t)

Ingår i:IJCAI International Joint Conference on Artificial Intelligence: International Joint Conferences on Artificial Intelligence, s. 3668-36769781956792034

Av författaren/redakt...: Eiter, Thomas; Geibinger, Tobia ...; Higuera, Nelson; Oetsch, Johannes

Om ämnet

Kungliga biblioteket hanterar dina personuppgifter i enlighet med EU:s dataskyddsförordning (2018), GDPR. Läs mer om hur det funkar här.
Så här hanterar KB dina uppgifter vid användning av denna tjänst.