Neurosymbolic AI in the Post-LLM Era: A Fusion Architecture of Hyperdimensional Computing and Probabilistic Soft Logic

This project explores the HDC-PSL-VSA-HDLM-AI architecture, integrating Hyperdimensional Computing (HDC), Probabilistic Soft Logic (PSL), Vector Symbolic Architecture (VSA), and Active Inference. It aims to build a post-LLM neurosymbolic AI engine, providing a new paradigm for localized forensic intelligence and addressing issues such as high resource consumption, hallucinations, and insufficient interpretability of traditional LLMs.

Although Large Language Models (LLMs) have transformed the AI field, they have limitations such as requiring massive computing resources, hallucination issues, and difficulty in interpretable reasoning. As a new paradigm, neurosymbolic AI combines the perceptual capabilities of neural networks with the reasoning capabilities of symbolic systems. The HDC-PSL-VSA-HDLM-AI project is a cutting-edge exploration of this trend.

Hyperdimensional Computing (HDC)

Inspired by the human brain, it uses extremely high-dimensional vectors (usually over 10,000 dimensions) to represent information, with fault tolerance and similarity preservation characteristics, laying the foundation for knowledge representation.

Probabilistic Soft Logic (PSL)

It handles uncertain knowledge, allowing fact confidence to range between 0 and 1, and can process fuzzy concepts, probabilistic reasoning, and conflicting evidence.

Vector Symbolic Architecture (VSA)

It provides algebraic operations such as binding and bundling, connects low-level perception with high-level cognition, and enables compositional reasoning.

Active Inference

Based on the Free Energy Principle, it unifies perception and action. Agents actively explore and verify hypotheses, reduce uncertainty, and achieve predictive processing.

Cryptographic Epistemology

It explores the nature of knowledge, trust, and proof in cryptography and distributed systems, providing a theoretical foundation for privacy-preserving intelligent systems (e.g., the PlausiDen deniability system).

Three-Tier Integration Architecture

• Neurosymbolic-Toolkit: The base layer, providing core primitives for HDC, VSA operations, and probabilistic reasoning;
• Shield: The control plane, responsible for coordination, security, and policy execution;
• Engine: The intelligent generation layer, executing complex reasoning and decision-making.

Rust Implementation

Rust is chosen to balance performance and security, with zero-cost abstractions and memory safety guarantees, making it suitable for large-scale hyperdimensional vector computing and parallel reasoning.

The project focuses on the localized forensic intelligence scenario: traditional cloud-based AI has data leakage risks, while offline solutions lack intelligence. This architecture can run efficient neurosymbolic reasoning on local devices, protecting privacy while providing intelligent analysis capabilities. The PlausiDen deniability system is an application example of this.

This project represents an important direction in the evolution of AI architectures, integrating neural and symbolic capabilities, and is expected to break through in interpretability, efficiency, and privacy protection. With the growth of edge computing and privacy demands, the value of localized, efficient, and interpretable AI architectures is becoming increasingly prominent. Its technical concepts will leave a mark on the development of AI.