← back to experiments

§ 2.1 — EXP-03 / Deep Learning

neural-network-optimization

A comprehensive C++ implementation of neural networks with optimization algorithms and evaluation on the Iris dataset.

Features

  • Custom neural network with configurable layers and activation functions.
  • Sigmoid, ReLU, and Softmax activation support.
  • Full evaluation pipeline: accuracy, precision, recall, F1-score, and confusion matrix.
  • K-fold cross-validation with stratified sampling.
  • Performance benchmarking with Google Benchmark.

Dependencies

  • Eigen3 for linear algebra.
  • Google Test for unit/integration tests.
  • Google Benchmark for performance measurements.
  • CMake as build system.

Cross-validation snapshot

  • Mean accuracy: 76.00% +/- 18.79%
  • Best fold: 100% accuracy
  • Worst fold: 53.33% accuracy