CubeMind Integration
CubeMind is a neuro-vector-symbolic reasoning system built on grilly. It combines hyperdimensional computing (HDC) with neural networks for structured symbolic reasoning on GPU.
Overview
CubeMind uses grilly's HDC operations (block codes, circular convolution, packed binary vectors) and neural network layers to build a reasoning system that:
- Represents knowledge as high-dimensional binary vectors
- Performs symbolic operations (binding, bundling, sequence encoding) on GPU
- Trains neural components with grilly's autograd and optimizers
Grilly Features Used by CubeMind
| Feature | CubeMind Usage |
|---|---|
| HDC packed ops | State representation as 10,240-bit hypervectors |
| Block-code circular conv | Sequence encoding for move histories |
| DisARM gradient estimator | Training through discrete block-code operations |
| Sanger GHA | Neurogenesis -- growing network capacity |
nn.Linear, nn.ReLU |
Neural prediction heads |
optim.AdamW |
Training neural components |
| VulkanTensor | GPU-resident state vectors |
C++ Backend Support
The C++ backend includes CubeMind-specific bindings in cpp/src/cubemind/ for high-performance dispatch of HDC operations that would be slow in pure Python.
Getting Started
pip install cubemind
CubeMind depends on grilly >= 0.5.0. See the CubeMind repository for documentation and examples.