NEURO-SYMBOLIC AI
Hybrid Intelligence is a Neuro-symbolic approach to AI. It combines the strengths of symbolic reasoning with the predictive power of neural networks to give:
- Full explainability: the explicit use of symbolic rules and logic enables a clear understanding of how the AI system arrives at its decisions, providing a trace of the reasoning process and precise and certain quantification of the influence of each feature on the output.
- Integration of structured and unstructured data that allows for a more comprehensive and holistic analysis, leveraging the advantages of both types of data.
- Flexibility and adaptability: to handle a wide range of complex problems and to adapt to different domains and tasks.
- Improved generalization and robustness through the incorporation of symbolic reasoning.
- Utilization of prior knowledge and expert insights into the learning process, leading to better outcomes and data efficiency.
HYBRID INTELLIGENCE MODEL
A Hybrid Intelligence model comprises two fully integrated parts:
- Explainable Neural Net (XNN): forms the neural substrate of a Hybrid Intelligence model and is a graph neural model that represents knowledge hierarchically. XNNs have a wide, modular architecture and provide powerful predictive performance with full transparency and explainability.
- Explanation Structure Model (ESM): The ESM is a hyper-graph model that encapsulates and integrates the XNN, and provisions and organises symbolic logic, reasoning, labels, and relationships. ESMs also incorporate data engineering and pre-processing pipelines.
Together, the XNN and ESM provide the full neuro-symbolic AI model. Hybrid Intelligence models are fully transparent and interpretable, and they communicate explanations concurrently with predictions.
INDUCTION
Hybrid Intelligence models are created directly from data using a unique model induction methodology that leverages several techniques including:
- Statistical and Deep learning techniques
- Information Theory
- Causal Analysis
- Symbolic Learning techniques
When appropriate, induction can also incorporate existing knowledge from black-box models and expert systems into the learning process.
MODEL OPTIMISATION & MAINTENANCE
Hybrid Intelligence leverages the transparency and modular structure of XNNs and ESMs to bring new and improved functionality to optimise and maintain models over their lifecycle.
- Surgical model optimisation for specific performance objectives
- Model compression with minimal effect on performance
- Selective, as needed, model re-training and/or re-structuring
- Targeted model re-training and re-structuring
- Lifetime traceability of model structure, logic and behaviour changes
REAL-TIME EXPLANATIONS & DECISIONS
When queried, a Hybrid Intelligence model produces and communicates its prediction and the complete and precise explanation data for that prediction concurrently. The availability of real-time explanations creates an entirely new interventional space to:
- Assess predictions in real-time to identify issues and sub-optimal behaviour
- Intervene directly and seamlessly to turn fit-to-data predictions into fit-for-purpose decisions
- Automate interventions in real-time workflows
- Selectively de-automate to bring a human into the loop where appropriate
- Curate explanations appropriately for different stakeholder needs
- Monitor and anticipate model behaviour dynamically
- Create accurate, verifiable and reproducible audit logs for each decision
