Our Team

Justin Diamond – PhD candidate in Computer Science (ML focus, finishing June).brings cutting-edge machine learning expertise.

Ziyu She – PhD student in AI with publications

Ryan Diamond, Advisor – guides financial strategy and business planning.

Floriane Le Floch, Advisor – Web3 consultant and Auxetic.ai founder

Company Name (if applicable)

Project details

Why This Project Is Needed

The limitations of current deep learning systems are well recognized: they require large amounts of data, struggle to reason over structured relationships, and are largely opaque to human understanding. These shortcomings pose critical challenges in domains where data is scarce, causal reasoning is essential, and transparency is non-negotiable — such as in scientific discovery, medicine, and autonomous systems. As artificial intelligence advances toward AGI and eventually ASI, these weaknesses must be addressed.

This project tackles the problem head-on by integrating symbolic logic into deep neural networks to enable higher-order reasoning, experiential learning, and small-data generalization. While neural networks excel at function approximation from high-dimensional data, they lack the ability to encode and apply known rules, causal structures, or logical abstractions. Conversely, symbolic systems are expressive, interpretable, and capable of structured reasoning, but are brittle and difficult to scale.

A truly general intelligence must combine these strengths. Neural-symbolic systems offer a principled way forward: a hybrid paradigm where logic guides learning, and learning refines logic. This project proposes to build and evaluate such systems within a high-fidelity simulation environment, using symbolic rules as both training priors and inference engines — implemented directly in the OpenCog Hyperon framework.

Simulation as an Ideal Testbed for Symbolic Reasoning

To ensure clarity, precision, and real-world applicability, this project will use physics and molecular simulations as a controlled environment to test neural-symbolic architectures. These domains are ideal for three reasons:

1. Causal Ground Truth and Formal Structure: Simulations are governed by well-understood physical laws, offering complete visibility into the causal relationships that underlie observed behaviors. Every entity, interaction, and transition is measurable and manipulable. This makes it possible to design experiments with known expected outcomes, enabling fine-grained testing of learned rules and inference accuracy.

2. Intervention and Counterfactual Analysis: Unlike most real-world data, simulation environments allow for controlled interventions. We can apply a force, change a molecule, or vary initial conditions to observe how a system responds. This is essential for evaluating symbolic inference, which relies not only on correlation but on the ability to generalize from structured, causal rules.

3. Scalable, Ab Initio Data Generation: The project does not require large pre-labeled datasets. Instead, data will be generated ab initio through scientific simulations (e.g., of molecules, forces, or particle interactions). This provides perfect labels and allows for the construction of both standard and edge-case scenarios — a critical advantage for small-data learning, curriculum design, and symbolic rule derivation.

By embedding symbolic reasoning into neural models operating over these simulations, we can rigorously evaluate the benefit of logic in improving generalization, robustness, and interpretability.

Technical Implementation Strategy

This project will explore two leading neural-symbolic paradigms:

1. PyNeuraLogic – a differentiable logic programming framework that enables the embedding of symbolic logic rules directly into neural network architectures, particularly Graph Neural Networks (GNNs). This will be used to encode logic such as “if atom A is bonded to B and B is an acid group, then A is reactive,” allowing logic-based supervision and constraint enforcement during learning.

2. Kolmogorov–Arnold Networks (KANs) – a recent class of networks that place learnable functions on the edges between nodes, offering both high accuracy and inherent interpretability. Their architecture lends itself well to modeling continuous physical processes alongside discrete symbolic flags, such as toggling functional forms based on rule satisfaction. KANs provide a natural bridge between continuous data and rule-based switching.

These architectures will be integrated into simulation-based reasoning tasks where symbolic rules are used in two ways:

• Experiential Logic: Derived from system interactions (e.g., from AIRIS or AIRIS-like behavior), where agents autonomously discover logic rules through exploration and pattern recognition. These rules will be used to guide future predictions or constrain possible states.

• Higher-Order Reasoning: Abstract rules supplied by experts (e.g., chemical valence rules, conservation laws, reaction heuristics). These will be encoded into the system manually or semi-automatically and enforced as constraints or priors in model inference.

Symbolic Representation via MeTTa and Atomspace

Central to the project is the use of the OpenCog Hyperon architecture, specifically its MeTTa language and Atomspace knowledge base. These tools allow symbolic knowledge (e.g., logic rules, facts, causal relationships) to be encoded in a dynamic hypergraph and used by reasoning engines to guide or verify model behavior.

In this system, simulation states will be converted into Atomspace structures — atoms representing particles, positions, energies, reactions, etc. Symbolic rules will be encoded as MeTTa expressions or pattern-matching templates. For example:

Such rules allow us to perform forward and backward reasoning within MeTTa. Rules can be activated during simulation episodes to derive expectations, filter impossible transitions, or backpropagate logic-based feedback into the neural learning loop.

The project will explore both tight and loose integration between neural models and the Atomspace:

• In tight coupling (e.g., PyNeuraLogic), logic rules are embedded into the architecture itself, influencing learning gradients.

• In loose coupling, Atomspace serves as an external knowledge base queried by the network for symbolic context, and used to check the consistency of model outputs.

This dual-mode integration allows flexibility, scalability, and modularity, aligning with Hyperon’s long-term cognitive architecture vision.

Learning from Small Data via Symbolic Rules

A key advantage of neural-symbolic models is that they can learn effectively from limited data by using rules as inductive biases. In scientific domains, where each datapoint can be expensive to obtain (e.g., molecular synthesis or clinical trials), data efficiency is critical.

Symbolic rules — such as known equations, relational constraints, or invariants — help guide the model toward plausible inferences even with little supervision. For instance, instead of needing to learn energy conservation from thousands of examples, the rule can be embedded and used to constrain the model’s predictions.

This project will explicitly quantify the benefit of logic in small-data regimes by comparing performance with and without symbolic priors, and evaluating generalization to out-of-distribution conditions.

Bridging Symbolic and Subsymbolic AI

The project is grounded in the recognition that true AGI requires both data-driven learning and structured reasoning. Neither alone is sufficient. This duality reflects human intelligence, where raw perception and abstract knowledge co-exist. Bridging this gap requires systems that:

• Learn from data but respect symbolic constraints.

• Reason abstractly using logic, but adapt via gradient-based learning.

• Integrate perceptual features and relational knowledge fluidly.

By implementing symbolic neural architectures in simulation-rich, causally-grounded environments, this project builds precisely such a bridge. It offers a concrete testbed for hybrid cognition — and contributes directly to the broader goals of SingularityNET and the Artificial Superintelligence Alliance: to engineer general-purpose intelligence that is interpretable, grounded, and robust.

Relevance to SingularityNET and AGI Objectives

This proposal aligns with SingularityNET’s mission to advance open, decentralized AGI in multiple ways:

• Hyperon Integration: By using MeTTa and Atomspace directly, the project contributes reusable components to the Hyperon/PRIMUS ecosystem, helping realize a general cognitive substrate.

• AGI-Ready Learning Systems: The architecture developed here — combining experiential learning, higher-order reasoning, and neural-symbolic control — represents a core step toward robust cognitive systems capable of reasoning and learning across domains.

• Neuro-Symbolic Synergy: As outlined by OpenCog and TrueAGI leadership, future AGI systems will require strong neural-symbolic integration. This project delivers precisely such a prototype, grounded in scientifically rigorous tasks.

• Scientific Impact: The proposed work lays a foundation for better AI in domains where safety, explainability, and correctness are essential — medicine, science, engineering — and where traditional AI approaches falter.

• Decentralized Ecosystem Alignment: By leveraging the Hetzerk platform for data generation and sharing, the project supports distributed, open, and collaborative experimentation in AI — furthering the goals of a decentralized AGI future.

Project Staging and Evaluation Approach

The project will progress through clearly delineated phases: surveying architectures and formalizing symbolic rule formats, developing working prototypes of neural-symbolic models, embedding symbolic knowledge into Atomspace, and conducting simulation-based evaluations comparing symbolic vs. non-symbolic learning. At each stage, results will be validated against control tasks and interpreted with respect to data efficiency, reasoning fidelity, and alignment with embedded logic.

The project will culminate in a reproducible, integrated MeTTa/Hyperon demonstration — showing an AI system reasoning over a structured simulation, guided by embedded rules, and learning more effectively as a result.

Links and references

Justin's google scholar:

https://scholar.google.com/citations?user=O-chxmAAAAAJ&hl=en

Justin's PhD evidence:

https://pharma.unibas.ch/de/personen/justin-diamond/

Hetzerk.com

Additional videos

https://youtu.be/jdyoPr5mNZ0?t=532

https://www.youtube.com/watch?v=F575gBjOCr4