New probabilistic neuro-symbolic framework boosts AGI reasoning with entropy

In a new arXiv preprint, researcher Zoran Majkic proposes a significant expansion of neuro-symbolic artificial intelligence by introducing a probabilistic layer on top of Belnap's Typed Intensional First-Order Logic (IFOL_B). This approach aims to overcome the limitations of pure neural systems—such as lack of interpretability and formal logical structure—by combining neural learning with symbolic reasoning. The key innovation is using Nilsson's probability structure to assign probabilities to currently unknown sentences, allowing the system to reason under uncertainty while maintaining logical consistency.

The framework introduces two types of symmetry transformations: a global symmetry that preserves the current knowledge database and logical deduction, and a local symmetry used for real-time decisions on concrete subproblems involving a strict subset of IFOL_B predicates. Both transformations leverage neural networks to compute the probability density function based on Shannon's maximum information entropy principle. This enables the AGI system to dynamically adjust its uncertainty estimates and make decisions that are both logically sound and probabilistically informed.

By grounding probabilistic reasoning in a formal logical system with self-reference capabilities, this work bridges the gap between statistical pattern recognition and rule-based deduction. The 32-page paper details how the integration allows neuro-symbolic robots to handle novel scenarios by computing probabilities for unobserved facts, while retaining the interpretability and structure of formal logic. This could lead to more robust and transparent AI systems capable of complex reasoning in dynamic environments.