Antonino Calderone

nuNN is a free, open-source machine learning library written in modern C++20 and released under the MIT License.

It provides compact and readable implementations of fundamental machine learning algorithms, with a focus on experimentation, education, and learning. The project is designed to make neural networks, reinforcement learning, and related algorithms understandable from source code, while still providing practical demos, tests, model serialization, and optional hardware acceleration.

The library includes:

Feedforward Neural Networks

Perceptron: single-neuron classifier with MSE or cross-entropy loss and momentum
MlpNN: classic fully connected multilayer perceptron with online SGD
MlpMatrixNN: Eigen-backed matrix MLP with mini-batch SGD, JSON serialization, and optional ArrayFire/OpenCL acceleration
RBF Network: radial basis function network with Gaussian centers and trainable output weights

Recurrent Neural Networks

VanillaRnn: Elman-style recurrent neural network with truncated BPTT
GRU: Gated Recurrent Unit with truncated BPTT
LSTM: Long Short-Term Memory network with truncated BPTT

Unsupervised and Generative Models

Hopfield Network: energy-based associative memory model
Autoencoder: symmetric encoder-decoder built on MlpMatrixNN
RBM / VAE support: restricted Boltzmann machine and variational autoencoder implementations

Convolutional and Attention Models

Conv1DLayer / MaxPool1DLayer / ConvNet: 1D convolutional pipeline with end-to-end backpropagation
LayerNorm / SelfAttentionLayer / TransformerBlock / MiniTransformer: decoder-only transformer components with causal self-attention and autoregressive generation

Reinforcement Learning

Q-learning: tabular off-policy temporal-difference learning
SARSA: tabular on-policy temporal-difference learning
DQN: Deep Q-Network with experience replay and frozen target network

Classical Machine Learning Utilities

K-means
PCA
Linear regression
Dataset and helper utilities, including an MNIST parser

Additional Features

Save and load complete model state, including JSON support
Export network topology to Graphviz DOT format
Optional OpenCL acceleration through ArrayFire, with automatic CPU fallback
Windows OpenCL runtime deployment helpers
OCR MNIST training dialog with backend selection, mini-batch training, persisted paths, and cost-convergence chart
Hundreds GoogleTest unit tests
Cross-platform support for Windows, Linux, and macOS

The project includes demo applications and tools for handwritten digit OCR, MNIST training, interactive Tic Tac Toe in console and Windows GUI versions, Hopfield pattern recall, maze and path-solving reinforcement learning, DQN grid-world learning, sine-wave prediction, character-level recurrent language modeling, transformer-based character generation, autoencoder reconstruction, RBF regression, convolutional sequence classification, and model conversion utilities.

Source Code: eantcal/nunn: Collection of Machine Learning Algorithms

Google Sites

Report abuse