[1Win4-08] Towards Neurosymbolic Reasoning that Mitigates Shortcut Reasoning
Keywords:logic programming, neurosymbolic AI, shortcut reasoning
Shortcut reasoning in neurosymbolic (NeSy) systems often leads to suboptimal generalization performance.
This experimental study investigates two types of shortcuts commonly observed in NeSy systems: \textit{constraint shortcut satisfaction} and \textit{reasoning shortcuts}.
The former arises when models trained with constraints, e.g., \((\neg P \vee Q)\), learn to predominantly satisfy one condition (e.g., \(\neg P\)) at the expense of learning to satisfy the other condition (\(Q\)).
The latter involves conceptual ambiguity, where models trained on confounded concepts struggle to differentiate between them, leading to reasoning errors.
Through a series of controlled experiments, we demonstrate the prevalence of these shortcuts in neural models and identify conditions under which they manifest.
These findings provide insights into improving the robustness of NeSy systems by addressing limitations in their reasoning processes.
This experimental study investigates two types of shortcuts commonly observed in NeSy systems: \textit{constraint shortcut satisfaction} and \textit{reasoning shortcuts}.
The former arises when models trained with constraints, e.g., \((\neg P \vee Q)\), learn to predominantly satisfy one condition (e.g., \(\neg P\)) at the expense of learning to satisfy the other condition (\(Q\)).
The latter involves conceptual ambiguity, where models trained on confounded concepts struggle to differentiate between them, leading to reasoning errors.
Through a series of controlled experiments, we demonstrate the prevalence of these shortcuts in neural models and identify conditions under which they manifest.
These findings provide insights into improving the robustness of NeSy systems by addressing limitations in their reasoning processes.
Authentication for paper PDF access
A password is required to view paper PDFs. If you are a registered participant, please log on the site from Participant Log In.
You could view the PDF with entering the PDF viewing password bellow.