Incorporating Human Knowledge into XNNs
Human Knowledge Incorporation connects explicit expert or organisational knowledge with the learnt structure of an Explainable Neural Network. Rules, constraints, taxonomies and causal models can be represented symbolically, reviewed and managed alongside data-derived behaviour.
Why data is not the only source of knowledge
Historical data reflects what happened, including its gaps, biases and obsolete practices. Decision systems also depend on knowledge that may not appear reliably in data: safety boundaries, policy obligations, domain classifications, procedures and cause-and-effect assumptions.
Hybrid Intelligence provides a structure in which these knowledge sources can be represented explicitly rather than left in documents or added through an unrelated downstream rule engine.
Knowledge that can be connected
The Hybrid Intelligence framework identifies several forms that can be represented within an Explanation Structure Model (ESM):
- expert-authored IF-THEN rules;
- business and policy constraints;
- taxonomies and ontologies;
- knowledge graphs;
- causal models;
- formulas and procedures;
- source-code logic; and
- review feedback on learnt rules.
Each item should carry provenance, ownership, status and version information. A human-authored rule is inspectable, but it is not automatically correct or authorised.
Rule injection and inspection
XNNs express learnt local behaviour through IF-THEN rules. HKI uses a compatible symbolic form:
IF condition THEN action or constraint
An expert can add a missing rule, review a learnt rule, refine a threshold or specify how a conflict should be handled. Decision tables and other visual formats can make the logic accessible to non-programmers.
The system should distinguish types of rules. A learnt predictive rule, a mandatory policy rule and an advisory expert rule have different authority and should not be merged into one undifferentiated list.
A controlled feedback loop
A typical workflow is:
- induce an initial XNN from data and available knowledge;
- expose relevant learnt rules and model behaviour;
- collect expert review with rationale and provenance;
- validate proposed additions or changes;
- update the ESM and, where required, retrain the XNN;
- compare the new version with its parent; and
- monitor how the rule behaves in operation.
This creates a traceable collaboration cycle. It should include approval, testing and rollback controls appropriate to the risk of the decision.
Hard and soft matching
Hybird Intelligence supports both hard and soft forms of rule matching. Hard matching requires exact conditions and suits deterministic requirements. Soft matching allows weighted or approximate matches and may suit noisy or uncertain domains.
The choice should be visible. A policy limit implemented as an exact constraint should not be presented as probabilistic guidance. Conversely, an approximate diagnostic pattern should not be misrepresented as an absolute rule.
Knowledge with limited data
Human rules can provide an initial structure when examples are scarce. This enables few-shot or, where a sufficiently complete rule base exists, zero-shot operation.
These are significant claims that require use-case-specific validation. The mechanism is that human-authored rules can initialise or constrain behaviour before extensive task-specific data is available. The quality and coverage of the result depend on the supplied knowledge and validation.
Taxonomies and causal structures
A taxonomy can connect specific features and categories to higher-level concepts. A causal graph can represent asserted or validated dependencies, such as one condition contributing to another.
The neuro-symbolic hypergraph links these structures to XNN objects. That enables explanations to move between a local model rule and a domain concept. It also allows what-if analysis to respect defined dependencies where the causal knowledge is appropriate.
Statistical interaction and causal relationship must remain distinct. An XNN may learn that two features jointly predict an outcome without proving that either causes it.
Conflict and lifecycle management
Human knowledge can conflict with data-derived behaviour, another policy or a later rule. A governed HKI process needs:
- rule priority and authority;
- conflict detection;
- effective dates;
- validation tests;
- decision logs;
- version comparison; and
- monitoring of affected queries.
Human control is meaningful only when changes are traceable and their consequences can be evaluated.
Key technical concepts explained
- Human Knowledge Incorporation (HKI): incorporation of explicit human or organisational knowledge into the Hybrid Intelligence structure.
- Rule provenance: evidence of where a rule came from, who owns it and why it is valid.
- Hard matching: exact evaluation of a rule condition.
- Soft matching: weighted or approximate rule evaluation.
- Knowledge authority: the status determining whether a rule is mandatory, advisory or learnt.
Simplified scenario
A clinical model learns a pattern suggesting that one group of cases can be discharged. A medical expert identifies a safety condition absent from the data and adds a rule requiring review when that condition is present. The rule is labelled as a safety constraint, tested against historical cases, approved, versioned and monitored separately from the predictive attribution.
Why it matters
HKI creates a practical bridge between model learning and accountable expertise. It allows organisations to represent what must be known, not only what the historical data happened to contain.