# Mind-Engine: Particle Neural Network and Synesthetic 3D Neuronal Cellular Automaton Built from Scratch

> Explore two original AI systems built entirely from scratch—Cognitron Particle Neural Network and Morpheus Synesthetic 3D Neuronal Cellular Automaton—without using pre-trained models, leveraging hyperdimensional computing encoding and WebGPU rendering technology.

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-04-29T23:42:26.000Z
- 最近活动: 2026-04-30T02:03:28.232Z
- 热度: 157.7
- 关键词: 粒子神经网络, 神经元胞自动机, 超维计算, 联觉生成, WebGPU, 无预训练模型, AI架构, 机器学习
- 页面链接: https://www.zingnex.cn/en/forum/thread/mind-engine-3d
- Canonical: https://www.zingnex.cn/forum/thread/mind-engine-3d
- Markdown 来源: floors_fallback

---

## Project Introduction: Mind-Engine—Particle Neural Network and Synesthetic 3D Neuronal Cellular Automaton Built from Scratch

The Mind-Engine project explores two original AI paths: Cognitron Particle Neural Network and Morpheus Synesthetic 3D Neuronal Cellular Automaton. Key features include: complete rejection of pre-trained models, use of hyperdimensional computing encoding and WebGPU rendering technology, and re-exploration of the original possibilities of AI architectures.

## Project Background and Core Concepts

Modern AI relies on pre-trained model libraries; most developers fine-tune on their basis but rarely question the dependency. The Mind-Engine team asks: What if we completely abandon pre-trained models and build AI systems from scratch? The core concept is to explore the original possibilities of AI architectures, with Cognitron (3D particle neurons) and Morpheus (cell rule-based growth) representing two different approaches.

## Cognitron Particle Neural Network Architecture and Implementation

### Architecture Overview
Cognitron implements a Particle Neural Network (PNN): neurons are particles in 3D space, and the connection graph is dynamically reconstructed based on spatial proximity.

#### Core Mechanisms
1. **Input Encoding**: Text is converted into 10,000-dimensional bipolar hypervectors via a Hyperdimensional Computing (HDC) encoder (based on hashing, no pre-trained embeddings);
2. **Particle Generation**: Hypervectors are randomly projected into 3D space to generate particles with position, velocity, mass, and charge;
3. **Forward Propagation**: Wave propagation mechanism—queries propagate 4-6 hops from input particles to calculate energy flow;
4. **Training**: Particle Gradient Descent (PGD) algorithm, which integrates Particle Swarm Optimization (PSO) and gradient descent.

#### Technical Implementation
- Core model implemented with NumPy, no ML library dependencies;
- Frontend: Next.js 16 + React Three Fiber + WebGPU rendering;
- Backend: FastAPI provides inference services, supporting real-time browser operation.

## Morpheus Synesthetic 3D Neuronal Cellular Automaton Design

### Architecture Overview
Morpheus extends Growing-NCA to 3D space, adding an audio frequency modality to achieve synesthetic generation of geometry, color, and sound.

#### Architecture Design
- 32×32×32 3D grid, each cell has 16-channel states (RGB, transparency, audio frequency, hidden states);
- Update process: 3D Sobel kernel perception → small MLP (16→96→16) state update → residual connection + 50% random masking → transparency pooling to determine survival;
- Synesthetic Binding: Auxiliary loss correlates audio frequencies with color hues (e.g., red corresponds to specific frequency bands).

#### Training and Inference
- Server-side training with PyTorch (no pre-trained weights), targets include shapes like spheres and spirals;
- Browser loads weights, users select target shapes, observe 3D growth, and experience synesthetic audio.

## Innovation Points and Research Value

### Innovation Points
- The combination of PNN in Cognitron (spatial neurons + PSO-gradient hybrid + HDC encoding + wave reasoning) has not been published in literature;
- Morpheus's audio-cell correspondence is the only new angle in the field of neuronal cellular automata (confirmed by a novelty-research agent).

#### Research Value
- Reactivates hyperdimensional computing (an ancient neuro-symbolic method) and combines it with modern deep learning;
- Provides new ideas for dynamic network topology;
- Opens up a new direction for multimodal AI (synesthetic generation).

## Use Cases and Interactive Demos

### Cognitron Use Cases
Users input text ideas → observe particles clustering in 3D space based on semantic relationships; after training, particles drift toward semantic neighbors; during queries, wave propagation lights up relevant particle clusters.

#### Morpheus Use Cases
Users select target shapes → observe cells growing into 3D structures from scratch; simultaneously play corresponding chords to experience visual-auditory synesthesia. Switching targets produces different shape and sound combinations.

## Limitations and Future Directions

### Limitations
- The convergence of the PGD algorithm is less stable than traditional optimizers (SGD fallback mechanism implemented);
- The expressive power of the hyperdimensional encoder needs verification;
- 3D NCA has high computational cost (default 32³ resolution).

#### Future Directions
- Expand the scale of particle neural networks;
- Explore the generation of more complex geometric shapes;
- Extend synesthesia to more modalities like touch and smell;
- Inspire more researchers to explore AI architectures beyond pre-trained models.

## Project Conclusion

Mind-Engine represents a bold exploration in the era dominated by pre-trained Transformers, reminding us that the possibilities of AI are far from exhausted. By building systems from scratch, the team not only created two usable tools but also reopened fundamental questions about the nature of neural networks and the forms of intelligent computing. This project is suitable for AI architecture enthusiasts and researchers, proving that original architecture design can still lead to unexpected discoveries even with limited resources.