The Hybrid Intelligence Induction Process

More than fitting parameters

Conventional model training usually begins with a predefined architecture and optimises its parameters. Hybrid Intelligence separates structural induction from subsequent weight training. Induction determines how the model should be organised: which features and interactions need modules, how their input spaces should be partitioned and which local rule structures should represent the behaviour.

This distinction is central to the lifecycle. Once an XNN structure has been induced, it may be retrained on compatible data without learning the structure again. If the problem changes materially, reinduction constructs a new structure.

Inputs to induction

The process can begin with data alone or with data plus symbolic knowledge. Potential knowledge inputs described in the framework include:

• causal models;
• knowledge graphs and ontologies;
• human-authored rules and procedures;
• model constraints;
• taxonomies;
• source-code logic; and
• knowledge derived from an existing model where an appropriate extraction process is available.

Every external input requires provenance and validation. The ability to incorporate a rule does not establish that it is complete, current or appropriate.

Stage 1: dataset analysis and preprocessing

The process first establishes the data schema and examines inputs and targets. It can calculate variable statistics, distributions, correlations and anomaly information. It determines how categorical, temporal, sequential or vector inputs will be transformed.

The transformations are compiled into an explainable preprocessing pipeline. Original features and their transformed representations remain linked, so a rule operating on an encoded parameter can still be traced back to the business field.

Stage 2: interaction discovery

Induction evaluates candidate relationships between features and targets. A module may represent one feature or a selected interaction. The objective is to capture material behaviour while controlling unnecessary complexity.

Interaction discovery is important for both prediction and explanation. If an effect depends on Age × Education, placing it in a named module makes the relationship inspectable rather than leaving it implicit across distributed parameters.

Stage 3: partition formation

Within each module, induction groups input combinations with similar output behaviour and separates dissimilar regions. Numerical and categorical features can be associated with bins that provide understandable boundaries or labels.

Partitions are organised hierarchically. Broader patterns can appear near the top and finer distinctions deeper in the structure. The hierarchy supports both local rule selection and comparison between neighbouring behaviours.

Stage 4: rule generation

Each partition receives an IF-THEN rule:

• the IF expression identifies the partition's conditions;
• the THEN expression defines a local linear or polynomial function.

Rules within one module are non-overlapping for a query, so one partition activates at a time. The local expression produces the module attribution used in the final prediction.

Stage 5: structural optimisation and training

Hybrid Intelligence prefers simpler, modular and relatively wide structures, with information-preserving symbolic mappings from inputs to outputs. The induced model structure is refined and its parameters are trained against the selected objective and evaluation criteria.

The output is a trainable XNN. It retains the learnt modules, partitions and rules together with the parameters required for later retraining.

Stage 6: neuro-symbolic packaging

The trained XNN is converted or prepared for serving and analysed into symbolic objects. Rules, bins, features and expressions receive canonical references and are connected within the neuro-symbolic hypergraph.

The ESM packages the explainable preprocessing, serving XNN, symbolic structure and metadata. This creates the runtime object used for explainable querying and lifecycle comparison.

Induction principles

Hybrid Intelligence follows several principles:

• group inputs with similar behaviour;
• decompose complexity into modules;
• prefer simple local forms where they are sufficient;
• favour structures that remain understandable;
• preserve symbolic lineage;
• incorporate all relevant knowledge sources; and
• retain the relationship between parent structures and derived rules.

These principles should be evaluated in implementation rather than repeated as absolute guarantees.

Why it matters

Induction determines what the model will be possible to inspect later. By making interaction selection, partition boundaries and local rules explicit, it creates the structural basis for explanations, targeted retraining and version-level comparison.