Welcome to The Brain! This is an experimental, high-performance Spiking Neural Network (SNN) engine designed to simulate biologically inspired learning processes directly on the GPU.
⚠️ Status: Work in Progress This project is in an early experimental phase. The code is built for research and rapid prototyping of neural architectures.
- GPU-First Architecture: Utilizes ILGPU to compute massive neural networks in parallel using Compute Shaders.
- ECS-Inspired Design: Efficient management of
NeuronStateandSynapseDatastructures to maximize performance for hundreds of thousands of neurons. - Custom Serialization: Robust, high-speed snapshotting via Newtonsoft.Json, featuring custom converters to handle C#
unsafe fixedbuffers efficiently. - Cross-Platform: Developed for Ubuntu and Windows using VS-Code.
- Self-Wiring Capabilities: Includes built-in mechanisms for structural plasticity and synapse candidate ranking.
To ensure smooth simulation and enough memory for the neural pool, The Brain requires:
- GPU: A CUDA-compatible graphics card (NVIDIA).
- VRAM: At least 8 GB of dedicated video memory.
- OS: Windows 10/11 or Ubuntu 20.04+ (tested with VS-Code).
🚀 Note: While the engine is highly optimized, the large input/output layers and the 400k-neuron hidden pool thrive on fast VRAM and high CUDA core counts.
- Language: C#
- Graphics/Compute: OpenTK (OpenGL Graphics), ILGPU (Compute Shaders)
- Serialization: Newtonsoft.Json (Custom Converters)
- GUI Framework: SummerGUI
- IDE: VS-Code
The Brain utilizes an optimized snapshot system to persist the neural state (including Iteration, Neurons, and Synapses). By implementing custom JSON converters, complex GPU memory structures are mapped into clean, human-readable JSON formats without losing data integrity.
- Refinement of Hebbian learning kernels.
- API documentation for external training dataset ingestion.
To run The Brain, you need the GPT-2 tokenizer files. Due to licensing and size, they are not included in this repository. Please follow these steps to set up your environment:
- Create a folder named
tokenizerin the project root directory. - Download the following two files from the official Hugging Face GPT-2 repository:
- Place both files into the
/tokenizerfolder you created.
The system will automatically detect and load these files upon the first build.
💡 Pro-Tip: If the files are missing, the engine will prompt you with a
FileNotFoundExceptionto ensure you have the necessary dependencies for the BPE-based tokenization.
As this project is experimental, I welcome constructive discussions and feedback regarding:
- GPU memory optimization.
- SNN topology research.
- Serialization efficiency for large-scale buffers.
Please note: Interfaces and data structures are subject to frequent changes.
Built with ❤️ and a lot of C#.