This project is a feed-forward neural network written in C++.
It's built completely from scratch, without the help of any 3rd-party library for matrix operations, differentiation operations or linear algebra.

The network achieves the following tesults:

Accuracy on unseen data of 90% on the Fashion MNIST dataset (multi-class classification) in 6 minutes with 8 threads,
config:

OUTPUT_LAYER_SIZE = 10
HIDDEN_LAYERS_NEURON_COUNT = { 128, 128 }
BATCH_SIZE = 200
NORMALIZATION_FACTOR = 255

Accuracy on unseen data of 100% on 16-bit XOR dataset (binary classification) in 3 minutes with 1 thread,
config:

OUTPUT_LAYER_SIZE = 2
HIDDEN_LAYERS_NEURON_COUNT = { 32, 32, 32, 32 }
BATCH_SIZE = 100
NORMALIZATION_FACTOR = 1

All of these datasets are located in the data folder.

1. Clone repo
2. Compile
   • cmake src/
   • make
3. Extract datasets
   • From data.zip to MLP/data, i.e. MLP/data/fashion_mnist_train_vectors.csv
4. Run the binary
   • ./main

The program will start working on the Fashion MNIST dataset with the configuration mentioned above. The dataset and configuration can be changed in main.cpp (the hyperparamter constants).