DOMAIN-GROUNDING OF NEURAL NETWORKS FOR SPATIOTEMPORAL REASONING

Kaushik Roy, Alessandro Oltramari, Chathurangi Shyalika, Yuxin Zi, Vignesh Narayanan, Amit P. Sheth

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Domain Grounding, Knowledge Graphs, Constrained Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We formulate the neural network as a dynamical system and formulate domain knowledge-dependent constraints over its internal structures during training. Our proposed methodology improves domain-grounding and downstream task performance.

Abstract: Neural Networks are powerful approximators for learning to reason from raw data (e.g., pixels, text) in spatio-temporal domains (e.g., traffic-scene understanding). However, several recent studies have shown that neural networks are prone to erroneous or sometimes absurd reasoning that lacks domain-grounding (e.g., adhering to intuitive physics and causality). Incorporating comprehensive symbolic representation for domain understanding as part of a consolidated architecture offers a promising solution. In this paper, we take a dynamical systems perspective of a neural network and its training process, and formulate domain knowledge-dependent constraints over its internal structures (parameters and inductive biases) during training. This is inspired by \textit{control barrier function}, a constraint specification method from control theory. In particular, we specify the domain knowledge using Knowledge Graphs in our approach. To demonstrate the effectiveness of our approach, we apply it to two benchmark datasets focused on spatiotemporal reasoning: CLEVRER and CLEVRER-Humans, both centered around the task of question answering. Furthermore, we propose novel ways to evaluate if domain-grounding is achieved using our method. Our results show that the proposed methodology improves domain-grounding and question-answering accuracy while endowing the model with enhanced interpretability - an interpretability score that specifies to which extent the domain constraints are followed or violated.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5685