Course outline for Neuro-Symbolic AI

Module 1: Assign complementary reasoning roles

• Symbolic AI: Represent explicit facts and domain rules for reviewable conclusions while recognizing that correct deduction still depends on true premises.
• Neural AI: Treat model-extracted skills and recommendations as candidate proposals whose evidence must be checked before acceptance into the decision pipeline.
• Rule-based reasoning: Apply reviewed role requirements to accepted premises and trace conclusions to rules without assuming deterministic execution guarantees factual correctness.

Practical outcome: Define a proposal/check boundary and a small reviewed rule set for synthetic role requirements.

Module 2: Connect facts to a hybrid pipeline

• Knowledge graphs: Preserve skill relationships and their supporting evidence in a bounded model rather than promoting every extracted relationship to fact.
• Ontologies: Define domain classes and relationships while keeping ontology entailment distinct from record validation and business-constraint enforcement.
• LLM + symbolic systems: Connect neural proposals to reviewed facts and symbolic checks through explicit boundaries that retain evidence and reject unsupported inputs.

Practical outcome: Map candidate skill facts into a bounded knowledge model, preserve their evidence, and distinguish ontology inference from validation.

Module 3: Check, abstain, and compare decisions

• Constraint checking: Validate record requirements and business restrictions separately from ontology inference, identifying invalid suggestions through explicit acceptance checks.
• Hybrid reasoning: Compare premise-linked hybrid decisions with model-only fixtures, reporting false acceptances, false rejections, abstentions, and review effort.
• Abstention and escalation: Withhold decisions on missing or contradictory evidence and assign review using the consequences of acceptance, rejection, and escalation.

Practical outcome: Produce premise-linked decisions and an accept/reject/escalate loss table for missing, conflicting, and invalid inputs.

Practical exercise

• Build a local hybrid decision pipeline from supplied extraction fixtures, synthetic profiles, and reviewed role rules.
• Accept supported valid facts; reject a constraint violation and escalate missing or contradictory evidence.
• Inject a false extracted premise and show why deterministic rules alone cannot guarantee a true conclusion.
• Compare model-only fixture recommendations and hybrid decisions on six labeled cases, reporting false acceptances, false rejections, abstentions, and review effort.