Neuro-symbolic weak supervision: theory and semantics

Area of Science:

• Artificial Intelligence
• Machine Learning
• Logic Programming

Background:

• Weak supervision presents challenges in reliability and semantic clarity for machine learning.
• Multi-instance partial label learning (MI-PLL) specifically struggles with ambiguous supervision and uncertain instance-label mappings.

Purpose of the Study:

• To propose a neuro-symbolic framework for structuring MI-PLL using inductive logic programming (ILP).
• To enhance the reliability and semantic clarity of machine learning models operating under weak supervision.

Main Methods:

• Integration of ILP to define a hypothesis space over label transitions.
• Formalization of per-instance classifier semantics within the ILP framework.
• Development of a relational scaffold for reasoning about weak supervision.

Main Results:

• Two inductive tasks were studied: inferring transition predicates (TPs) and instance-level classifier assignments.
• The formal semantics aids in constraint specification, consistency checking, and diagnosing semantic failure modes.
• The framework provides a structured approach to address challenges in MI-PLL.

Conclusions:

• The proposed neuro-symbolic framework offers a structured method for MI-PLL under weak supervision.
• Formal semantics improve the understanding and diagnosis of model behavior in complex labeling scenarios.
• This approach contributes to more robust and reliable AI systems, particularly for safety-critical applications.