NeuroxAI

GPU-Accelerated Neuromorphic Computing Platform

Biologically-inspired spiking neural networks targeting 1-10M neurons with real-time performance

Overview • Installation • Usage • Architecture • Performance • Documentation

---

Overview

NeuroxAI is a high-performance neuromorphic computing platform that simulates biologically-realistic spiking neural networks. Built in Rust with CUDA acceleration, it implements state-of-the-art learning algorithms and cognitive architectures inspired by the human brain.

Key Features

• 🚀 GPU Acceleration — CUDA-optimized kernels for RTX GPUs, targeting 1-10M neurons
• 🧠 Biological Accuracy — LIF neurons, STDP learning, realistic synaptic dynamics
• 🔬 Advanced Plasticity — Triplet STDP, BCM metaplasticity, homeostatic regulation
• 🌊 Neural Oscillations — Theta-gamma coupling, sleep consolidation cycles
• 💬 Language Processing — Dual-stream architecture (ventral/dorsal pathways)
• 🗣️ Conversational AI — Sequence generation, context windows, paraphrase detection
• 🎯 Cognitive Architecture — Working memory, attention, hippocampal memory systems
• ⚡ Neuromodulation — Dopamine, serotonin, norepinephrine dynamics
• 📊 Training System — JSON-based supervised learning with emotional patterns

Requirements

• Rust 1.75 or later
• CUDA (optional): For GPU acceleration, install CUDA Toolkit 12.6+ and use an NVIDIA GPU (Compute Capability 7.0+, RTX recommended). Without CUDA, build with --no-default-features for a CPU-only binary (brain runs, no MNIST benchmark/training).
• Windows/Linux (macOS not supported when using CUDA)

Installation

# Clone the repository
git clone https://github.com/TheRemyyy/neurox-ai.git
cd neurox-ai

# Build in release mode (with CUDA, default)
cargo build --release

# Build without CUDA (no NVIDIA stack required)
cargo build --release --no-default-features

Usage

System Information

cargo run -- info

Displays GPU capabilities and system configuration.

Interactive Chat Mode

cargo run -- chat

Starts an interactive session with the neuromorphic brain. Available commands:

Command	Description
/train <file>	Train from file (.txt for skip-gram, .json for supervised)
/vocab	Display learned vocabulary
/sleep	Run memory consolidation cycle
/debug	Show internal brain state
/help	List all commands
exit	Quit the application

Project Structure

neurox-ai/
├── src/
│   ├── main.rs              # CLI entry point
│   ├── lib.rs               # Library exports
│   ├── brain/               # Whole-brain architecture
│   ├── cortex/              # Cortical processing (V1, working memory, etc.)
│   ├── cuda/                # GPU kernels and context
│   ├── language/            # Dual-stream language + sequence generation
│   ├── learning/            # STDP, metaplasticity, homeostasis
│   ├── memory/              # Hippocampal memory systems
│   ├── semantics/           # Embeddings, paraphrase detection, intent clustering
│   ├── affect/              # Emotional states and transitions
│   ├── neuron/              # Neuron models (LIF, dendritic, interneurons)
│   ├── neuromodulation/     # Dopamine, serotonin, norepinephrine
│   ├── oscillations/        # Neural rhythms and coupling
│   ├── synapse/             # Synaptic dynamics, vesicle cycles
│   └── ...
├── data/                    # Training data
└── tests/                   # Test suite

Architecture Highlights

Neuron Models

• Leaky Integrate-and-Fire (LIF) with adaptive threshold
• Dendritic compartment models
• Interneuron subtypes (PV, SST, VIP)

Learning Mechanisms

• Triplet STDP with eligibility traces
• BCM metaplasticity for dynamic thresholds
• Reward-modulated learning (R-STDP)
• Homeostatic synaptic scaling

Memory Systems

• Hippocampal pattern separation (DG) and completion (CA3/CA1)
• Working memory with capacity limits
• Sleep-dependent consolidation

Conversational AI

• Sequence Generator — Beam search with temperature control
• Context Window — Maintains conversation history for coherent dialogue
• Paraphrase Detection — Recognizes semantically similar inputs
• Intent Clustering — Groups related intents with semantic similarity
• Emotional Responses — Sentiment patterns and emotion transitions

Performance

Optimized for NVIDIA RTX GPUs with:

• Sparse matrix operations (cuSPARSE)
• Parallel neuron updates
• Efficient spike propagation
• Memory-optimized data layouts

Documentation

Comprehensive documentation is available in the docs/ directory:

• General Overview — Philosophy and high-level goals
• System Architecture — Detailed data flow and brain loops
• Cognitive Modules — Theory of Mind, Inner Dialogue, and 2025 upgrades
• Changelog — Version history and updates

Contributing

Contributions are welcome! Whether you're fixing a bug, improving biological accuracy, or optimizing CUDA kernels, please read our Contributing Guidelines first.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

Inspired by computational neuroscience research and modern deep learning frameworks. Special thanks to the Rust and CUDA communities.

---

_{Built with ❤️ and Rust}