Glossary of Terms

Course 7 - Beyond Deep Learning Gen AI With RGMs through Active Inference

A

Active Inference – A predictive AI framework based on the Free Energy Principle, where Intelligent Agents minimize uncertainty by continuously updating their internal models through perception, learning, and action.

Active Learning – A learning strategy where AI prioritizes gathering the most informative data points to refine its models, reducing data requirements.

Active Selection – The process of AI choosing the best action from multiple possibilities based on predicted future outcomes.

Adaptive AI – AI that continuously adjusts to real-world changes by dynamically updating its models instead of relying on static training data.

B

Beyond Generative AI – The shift from traditional Generative AI models (like ChatGPT and DALL-E) to more advanced frameworks like Renormalizing Generative Models (RGMs) that integrate reasoning, learning, and decision-making.

Black Box Problem – A limitation of deep learning where AI models produce outputs without transparent, interpretable decision-making processes.

C

Causal Learning – AI’s ability to understand cause-and-effect relationships rather than just recognizing statistical correlations in data.

Conceptual Modeling – The ability of AI to understand and represent abstract concepts and relationships between data points rather than just recognizing patterns.

Coarse-Graining – A process in which AI reduces complexity by simplifying data into hierarchical structures while preserving essential relationships.

Computable Context – AI’s ability to interpret and act upon contextualized digital and physical information using structured data.

D

Deep Learning – A subset of machine learning that relies on large-scale neural networks, but struggles with adaptability, efficiency, and interpretability compared to RGMs.

Discrete State-Space Representation – A categorization method where AI assigns distinct labels to data states (e.g., "hot" vs. "cold"), improving interpretability.

Dynamic Adaptation – The ability of AI to continuously refine its model based on real-world interactions and changing conditions.

E

Energy-Efficient AI – AI systems designed to operate with minimal energy consumption by leveraging renormalization and Active Inference to optimize processing.

Explainable AI (XAI) – AI models designed with transparency, ensuring their decision-making process is interpretable and trustworthy.

F

Free Energy Principle – A foundational neuroscience and AI concept, developed by Karl Friston, which states that intelligent systems seek to minimize uncertainty by continuously refining their models.

G

Generalization in AI – The ability of AI to apply learned knowledge across different domains and tasks without requiring extensive retraining.

H

Hierarchical Data Processing – AI’s capability to analyze data at multiple levels of abstraction, from fine-grained details to broad contextual understanding.

Holon Architecture – A nested ecosystem where each Intelligent Agent functions both autonomously and as part of a larger intelligent system.

I

Intelligent Agent – An autonomous system that perceives its environment, updates its beliefs, plans and takes actions using Active Inference principles.

Interpretability in AI – The ability to understand and explain how AI systems arrive at their conclusions, a major advantage of RGMs over traditional deep learning models.

Interoperability – The seamless integration of AI models, data, and digital environments, allowing Intelligent Agents to work across multiple domains.

M

Markov Decision Process (MDP) in Reinforcement Learning – A reinforcement learning framework where AI makes decisions based only on the current state, without considering past events.

Markov Decision Process (MDP) in RGMs - MDPs extend beyond single-state decision-making by incorporating paths as latent variables, meaning that decisions consider entire sequences of past states and actions, not just the present. This path-based approach enhances AI’s ability to improve decision-making, understand complex behaviors and dependencies, infer optimal long-term strategies, and adapt dynamically.

Multiscale Learning – AI’s ability to learn across different levels of abstraction, from micro-level details to high-level strategic planning.

N

Neural-Symbolic AI – A hybrid AI approach that integrates deep learning (neural networks) with symbolic reasoning to improve adaptability and understanding.

P

Perception and Planning Integration – The unification of perception (recognizing patterns) and planning (choosing actions) within a single AI model, a core capability of RGMs.

Predictive Coding – A process where AI generates expectations about incoming data and refines its understanding based on whether those expectations align with reality.

Path-Based Planning – AI’s method of predicting and selecting optimal decision-making paths based on learned conceptual models.

R

Real-Time Learning – AI’s ability to continuously update its knowledge and adjust decisions based on live data rather than relying solely on static pre-trained models.

Recursive Learning – A learning approach where AI iteratively updates its understanding based on repeated exposure to data, refining its internal representations over time.

Renormalizing Generative Models (RGMs) – A new class of AI models that apply renormalization techniques to efficiently process hierarchical data, reducingcomputational costs and improving adaptability through a single unified framework.

Reasoning AI – AI that not only recognizes patterns but also understands the conceptual relationships between them, enabling deeper insights and decision-making.

S

SAFE AI – A framework ensuring AI systems are Secure, Accountable, Fair, and Explainable, addressing common issues in traditional AI models.

Scale-Free Modeling – A key advantage of RGMs where AI can process data across multiple scales without losing accuracy, from microscopic details to large-scale planning.

Self-Organizing AI – AI that autonomously restructures its internal representations to optimize learning and decision-making.

Spatial Web – The evolution of the internet that integrates AI, blockchain, and immersive technologies, allowing Intelligent Agents to interact in digital and physical environments.

State-Space Representation – A mathematical framework that represents all possible states or configurations of a system, allowing RGMs to model complex environments dynamically.

T

Task Generalization – AI’s ability to adapt its learning to new, unseen tasks without extensive retraining.

Trustworthy AI – AI systems that prioritize reliability, security, and ethical alignment to build user and institutional trust.