Examples Gallery

Examples Gallery

The examples are part of the documentation. They are small enough to read, but each one isolates a specific learning idea.

Supervised Basics

Example	Main model	What it demonstrates
and_test	Perceptron	Linear separation and the perceptron update
xor_test	MlpNN	Why hidden layers are needed for nonlinear separation
mlp_matrix_xor	MlpMatrixNN	The same XOR idea through matrix-based training
linear_regression_demo	LinearRegression	OLS vs gradient descent on linear data

Use these first when checking a build or reading the library for the first time. They are intentionally tiny: if these do not work, larger demos will only hide the problem.

MNIST and OCR

Example	Main model	What it demonstrates
mnist_test	MlpMatrixNN / MlpNN	Training and evaluating a digit classifier
ocr_test	MlpMatrixNN / MlpNN	Interactive drawing, model loading, and MNIST training from GUI
net2json	conversion tool	Migration from legacy .net model files to JSON

mnist_test is the controlled experiment: fixed dataset, known train/test split, reproducible metrics. ocr_test is the user-facing experiment: it shows what happens when real drawings differ from the MNIST distribution.

Recurrent Models

Example	Main model	What it demonstrates
rnn_sine	RNN / GRU / LSTM	Time-series prediction and hidden state
rnn_char	RNN / GRU / LSTM	Character-level next-symbol prediction
rnn_adding	RNN / GRU / LSTM	Memory over a sequence

These examples are useful for comparing vanilla recurrence with gated recurrence. The important observation is not only final loss, but how quickly each model learns and whether it keeps information over longer spans.

Convolution and Transformer

Example	Main model	What it demonstrates
cnn_seq	ConvNet	1D local filters, max pooling, and MLP head
transformer_char	MiniTransformer	Decoder-only self-attention and autoregressive generation

cnn_seq is deliberately 1D so the sliding-window mechanics remain easy to inspect. transformer_char avoids external tokenizers and keeps the vocabulary small, making the attention path easier to debug.

Unsupervised and Representation Learning

Example	Main model	What it demonstrates
hopfield_test	HopfieldNN	Associative recall from noisy or incomplete patterns
kmeans_demo	KMeans	Clustering by nearest centroid
pca_demo	Pca	Linear dimensionality reduction
ae_demo	Autoencoder	Reconstruction through a bottleneck
rbm_demo	Rbm	Probabilistic reconstruction with Contrastive Divergence
vae_demo	Vae	Smooth latent space, reconstruction, and sampling
rbf_demo	Rbf	Distance-based hidden units for function approximation
som_demo	Som	Topological organization of prototypes

These examples show different meanings of learning without labels: grouping, projection, reconstruction, memory, and self-organization.

Reinforcement Learning

Example	Main model	What it demonstrates
maze	Q-learning / SARSA	Tabular value learning in a grid world
path_finder	Q-learning / SARSA	Path solving on graph-like environments
dqn_maze	Dqn	Neural Q-learning with replay buffer and target network

The maze examples are best read with the reward function open. In reinforcement learning, the reward design is often as important as the update formula.

Games and GUI Demos

Example	Main model	What it demonstrates
tictactoe	MlpNN	Console interaction and game-state evaluation
winttt	MlpNN	Windows GUI version of Tic Tac Toe
ocr_test	MlpMatrixNN / MlpNN	Windows GUI OCR and model diagnostics

The GUI demos are useful for packaging and runtime checks because they exercise installed resources, model discovery, icons, launcher behavior, and optional OpenCL runtime availability.

Suggested Reading Order

1. and_test
2. xor_test
3. linear_regression_demo
4. mnist_test
5. ocr_test
6. rnn_sine
7. cnn_seq
8. transformer_char
9. maze
10. dqn_maze

This path moves from simple supervised learning to sequence modeling, structured inputs, attention, and reinforcement learning.