How dkconnect/kumo-no-su works
Kumo No Su is a NumPy neural-network project whose distinguishing feature is learnable polynomial curves on connections, with nodes summing incoming edge outputs. The main demonstrated workflow trains a model on input/target data and extracts a symbolic formula; the repository also contains an MNIST training route and learned-edge visualization. The tree shows two implementations: the newer learning-2 model used by the root example, and the separate kumo package. Some subsystem links are omitted because their source was not sampled.
Main entry points: main.py Formula demo · kumo_network.py Kumo network · kumo_layer.py Polynomial layer · mnist_main.py MNIST entry
Components 14
Polynomial model
- Kumo network
learning-2/kumo_network.py - Polynomial layer
learning-2/kumo_layer.py - Polynomial basis
learning-2/kumo_layer.py
Training workflows
- Formula demo
main.py - MNIST entry
learning-2/mnist_main.py - Dataset loading
learning-2/dataset_loader.py - Pattern training example
experiments/train_network.py
Model inspection
- Formula extraction
learning-2/symbolic_exporter.py - Edge visualization
learning-2/visualize_edges.py - MNIST visualization
learning-2/visualize_mnist.py
Earlier implementation
- Kumo network
kumo/network.py - Polynomial layer
kumo/layer.py
Other components
- Data / caller
- Predictions / formula
Connections 14
- Data / caller → Formula demo: supplies data
- Formula demo → Kumo network: constructs and fits
- Kumo network → Polynomial layer: forwards / backpropagates
- Polynomial layer → Polynomial basis: expands inputs
- Formula demo → Formula extraction: extracts formula
- Formula demo → Predictions / formula: prints formula
- Kumo network → Predictions / formula: produces predictions
- MNIST entry → Dataset loading: loads data
- MNIST entry → Kumo network: trains model
- Edge visualization → Polynomial layer: reads coefficients
- MNIST visualization → Data / caller: shows results
- Data / caller → Pattern training example: supplies data
- Pattern training example → Kumo network: trains / predicts
- Kumo network → Polynomial layer: forwards / backpropagates