Implementation Map

Implementation Map

Use this page when moving from a concept to the code. Each row links the public contract, implementation, focused test, and smallest relevant example. Header-only templates show “header” in place of a separate source file.

Core training primitives

Topic	Public contract / implementation	Tests or use
numeric vector	nu_vector.h · nu_vector.cc	test_vector.cc
activations	nu_activation.h	test_mlpnn_activations.cc
cost functions	nu_costfuncs.h · nu_costfuncs.cc	test_costfuncs.cc
generic epoch trainer	nu_trainer.h	test_trainer.cc
neuron state	nu_neuron.h	used by Perceptron and MlpNN

The smallest complete read is: activation and vector helpers, Perceptron::feedForward(), Perceptron::backPropagate(), its test, then and_test.

Feedforward networks

Model	Header	Implementation	Tests	Example
Perceptron	nu_perceptron.h	nu_perceptron.cc	test_perceptron.cc	and_test.cc
MlpNN	nu_mlpnn.h	nu_mlpnn.cc	test_mlpnn.cc	xor_test.cc
MlpMatrixNN	nu_mlpmatrixnn.h	nu_mlpmatrixnn.cc	test_mlpmatrixnn.cc	mnist_test.cc
base model abstraction	nu_nn_model.h	nu_nn_model.cc	persistence tests above	model loaders

To compare the two MLP implementations, trace these operations in both source files:

constructor -> weight initialization
setInputVector -> feedForward
output delta -> hidden deltas
weight/bias update
calcMSE / calcCrossEntropy
toJson -> loadJson

MlpMatrixNN additionally contains batch matrix assembly, backend selection, Adam state, and host/device synchronization.

Recurrent networks

Model	Header	Implementation	Tests	Examples
VanillaRnn	nu_rnn.h	nu_rnn.cc	test_rnn.cc	rnn_sine, rnn_char
Gru	nu_gru.h	nu_gru.cc	test_gru.cc	same shared-interface demos
Lstm	nu_lstm.h	nu_lstm.cc	test_lstm.cc	rnn_adding

Read step() before bptt(). In the backward path, identify saved per-time-step intermediates, truncation boundaries, gradient clipping, and when state is advanced or reset.

Convolution and transformer

Component	Header	Implementation	Tests	Example
Conv1DLayer / MaxPool1DLayer	nu_conv.h	nu_conv.cc	test_cnn.cc	cnn_seq.cc
ConvNet	nu_convnet.h	nu_convnet.cc	test_cnn.cc	same
LayerNorm / SelfAttentionLayer / TransformerBlock / MiniTransformer	nu_transformer.h	nu_transformer.cc	test_transformer.cc	transformer_char.cc

For convolution, trace channel-major flattening, im2col, activation, saved max indices, and reverse gradient routing. For attention, trace shapes through Q/K/V projection, causal masking, row softmax, head concatenation, residuals, and output logits.

Classical, associative, and representation models

Model	Header	Implementation	Tests	Example
LinearRegression	nu_linear_regression.h	nu_linear_regression.cc	test_linear_regression.cc	demo
KMeans	nu_kmeans.h	nu_kmeans.cc	test_kmeans.cc	demo
Pca	nu_pca.h	nu_pca.cc	test_pca.cc	demo
HopfieldNN	nu_hopfieldnn.h	nu_hopfieldnn.cc	MLP/persistence suite and demo assertions	hopfield_test.cc
Autoencoder	nu_autoencoder.h	nu_autoencoder.cc	test_autoencoder.cc	demo
Rbf	nu_rbf.h	nu_rbf.cc	test_rbf.cc	demo
Rbm	nu_rbm.h	nu_rbm.cc	test_rbm.cc	demo
Vae	nu_vae.h	nu_vae.cc	test_vae.cc	demo
Som	nu_som.h	nu_som.cc	test_som.cc	demo

For these models, start with fit() or train() and identify exactly which parameters are learned. For example, Rbf::fitCenters() fixes centers and widths before train() changes only output weights.

Reinforcement learning

Component	Header / implementation	Tests	Example
Q-learning template	nu_qlearn.h	test_rl.cc	maze.cc
SARSA template	nu_sarsa.h	test_rl.cc	maze.cc
epsilon-greedy policy	nu_e_greedy_policy.h	test_rl.cc	maze
softmax policy	nu_softmax_policy.h	test_rl.cc	maze
graph Q-learning	nu_qlgraph.h · nu_qlgraph.cc	test_qmatrix.cc	path_finder.cc
replay buffer	nu_replay_buffer.h	test_dqn.cc	DQN maze
DQN	nu_dqn.h · nu_dqn.cc	test_dqn.cc	dqn_maze.cc

The tabular learners are templates, so their update code lives in the headers. Read the example's Agent contract before the learner: state transition and reward semantics belong to the environment, not to QLearn or Sarsa.

MNIST, persistence, and tools

Area	Source	Related executable/test
IDX parsing and normalized digit vectors	mnist.h · mnist.cc	mnist_test.cc
OCR training and recognition	ocr_test.cpp	ocr_test on Windows
OCR runtime fallback	ocr_launcher.cpp	packaged launcher
legacy-to-JSON conversion	net2json.cc	net2json
topology extraction and Graphviz output	nunn_topo.cc	nunn_topo
persistence behavior	each model source	test_mlpnn.cc, test_mlpmatrixnn.cc, test_perceptron.cc

A productive reading order

For any algorithm:

1. Read the public header until you can state input shape, output shape, learned parameters, and error conditions.
2. Open the smallest example and trace construction, training, and evaluation.
3. Follow one forward pass in the source.
4. Follow one update backward, writing down every saved intermediate.
5. Read tests for invalid shapes, numerical behavior, and round trips.
6. Change one example parameter and predict the result before running it.

Keep reading

Return to Theory Notes for equations, or Examples Gallery for executable-centered study paths.