VoodooNet is a non-iterative, zero-train neural architecture that replaces the "thermodynamic cooling" of Gradient Descent with instantaneous Galactic Expansion.
By projecting data into high-dimensional manifolds and solving for weights using the Moore-Penrose pseudoinverse, VoodooNet achieves superior accuracy to standard SGD in a fraction of the time.
| Dataset | Method | Hidden ( |
Accuracy | Training Time |
|---|---|---|---|---|
| MNIST | VoodooNet | 4,000 | 98.10% | < 1s |
| Fashion-MNIST | VoodooNet | 4,000 | 86.63% | ~94s |
| Fashion-MNIST | SGD (10 Epochs) | 64 | 84.41% | ~9s |
Note: VoodooNet hits "Hyper-Galactic" accuracy on Fashion-MNIST without a single step of backpropagation.
Unlike traditional networks that "learn" features slowly, VoodooNet treats the hidden layer as a Magic Hat.
-
The Expansion: Inputs are projected into a massive hidden space (
$d=500$ to$4000$ ) using high-entropy random weights. - The Discovery: We skip training. Instead, we use a closed-form analytic solution (Pseudoinverse) to discover the output weights in a single matrix operation.
- The Scaling: Accuracy scales near-logarithmically with dimensionality ($Accuracy \propto \log(d)$).
git clone https://github.com
cd VoodooNet
pip install -r requirements.txt
python run_galactic_mnist.py --dims 2000 --dataset fashion