What is Hybrid Intelligence?
Hybrid Intelligence is UMNAI's neuro-symbolic framework for building decision models whose predictions can be examined through explicit modules, partitions, rules, attributions and symbolic relationships.
A framework for decisions that must be understood
Many enterprise AI systems are assembled from disconnected parts: a preprocessing pipeline, a predictive model, a rule engine, a monitoring layer and a reporting interface. Each may be individually useful, but it can be difficult to establish a precise line from the original input to the final decision and its explanation.
Hybrid Intelligence is designed around that line of sight. It treats explanation as part of the model architecture and represents learnt knowledge in forms that can be addressed symbolically. The objective is to make prediction, explanation, human knowledge and lifecycle evidence work as parts of one structured system.
Three connected components
The first component is the Explainable Neural Network (XNN). An XNN decomposes a predictive task into modules. A module represents one feature or a feature interaction. Within each module, the input space is divided into non-overlapping partitions. Each partition is associated with an IF, THEN rule: the IF expression defines when the partition applies, and the THEN expression calculates the module's contribution.
The second component is the neuro-symbolic hypergraph. Features, categorical values, partitions, rules and other concepts receive unique references. The hypergraph connects these objects to each other and to external symbolic knowledge such as taxonomies, constraints or causal structures. Because a hypergraph can represent multi-way relationships, it can express interactions that involve several features or concepts.
The third component is the Explanation Structure Model (ESM). The ESM contains or references the explainable preprocessing pipeline, one or more XNNs, the neuro-symbolic structure and dataset metadata. It provides the context needed to generate predictions and organise explanations for different uses.
How a query moves through the framework
A query begins as a row or batch of input data. The ESM's preprocessing pipeline applies the same documented transformations used during model creation, such as categorical, time or vector encoding. The transformed values enter the XNN.
Each relevant XNN module evaluates the input and activates one partition. The active rule's THEN expression produces an attribution: a signed contribution to the output. Module contributions are combined with a background value and passed through the appropriate output function for regression or classification.
At the same time, the system can retain the active modules, partitions, rules, feature and interaction attributions, and the symbolic path associated with the result. The ESM can use those objects to create different views: a concise feature rollup, a detailed rule trace, a module view or a broader dependency explanation.
Learning from data and knowledge
Hybrid Intelligence models can be induced from data and can also incorporate symbolic knowledge. Potential sources described in the framework include expert-authored rules, taxonomies, ontologies, causal models, knowledge graphs, procedures and source-code logic.
Human Knowledge Injection does not make all supplied knowledge automatically correct. Rules still require provenance, validation, conflict handling and lifecycle control. Its importance is architectural: expert knowledge does not have to remain outside the model as an informal document or an unrelated rule engine. It can be represented as part of the explainable structure and reviewed alongside learnt rules.
Explanation at several levels
Hybrid Intelligence uses the same underlying attribution logic at different scopes:
- Local: why a particular query produced its result.
- Module-level: how one feature or interaction behaves across its partitions.
- Global: which modules, features, rules or activation patterns matter across a dataset or population.
This shared basis reduces the gap between a user-facing explanation and a technical model audit. The presentation changes, but the explanation can still be linked to the same model-native components.
Lifecycle and verification
The framework distinguishes induction, training, retraining and reinduction. Induction learns model structure, including modules, partitions and interactions. Training optimises parameters within that structure. Retraining updates an existing structure where appropriate. Reinduction constructs a new structure when the data schema, targets or underlying assumptions change materially.
At runtime, model versions and query outputs can be associated with Unique Verification Codes (UVCs). These cryptographic hashes are designed as tamper-evident fingerprints over selected model or query information. A matching recomputed UVC supports an integrity check; it does not by itself establish that the model is accurate, fair or compliant.
Why it matters
Hybrid Intelligence provides a common structure for people who normally see different fragments of an AI system. Data scientists can inspect model behaviour, domain experts can review rules and concepts, operational users can receive concise explanations, and assurance teams can trace outputs to versions and activation paths.