Perceptron Iris Classifier - From Scratch

This project demonstrates how a perceptron - one of the simplest neural models - can be built entirely from scratch in C++.
No machine learning libraries, no high-level frameworks: just basic data structures, loops, and arithmetic.

The program classifies samples from the Iris dataset into two categories by directly implementing:

• manual dataset parsing,
• handcrafted train/test splitting,
• weight initialization with basic random numbers,
• perceptron learning rule with explicit updates,
• accuracy evaluation without external tools.

---

• Raw Data Handling

  • CSV parsing implemented line by line using streams.
  • Features stored in a simple Input struct with std::vector<double> for attributes and an integer label.

• Perceptron from First Principles

  • Weights are just a std::vector<double>.
  • Predictions use a dot product + bias.
  • Error drives updates according to the original Rosenblatt learning rule.

• Train/Test Split Without Libraries

  • Random shuffling and slicing of vectors.
  • Implemented directly with STL algorithms (std::shuffle) instead of scikit-learn–style utilities.

• Accuracy Check

  • Counting correct predictions with explicit loops.
  • No metrics libraries - only integer counters and division.

• Visualisation of Learning

  • Saving necessary data to csv
  • Python script to process and plot the data
  • Plotting in Matplotlib

📸 Example Screenshots

• Charts plotted with Matplotlib showig impoving accuracy over each epoch

  

---

---

• Accuracy after every epoch and final accuracy of the model

---

• Learning process info (debug flag = true)

---

• Adjusting learning paramters (eta, epochs) influences the model performacne, for example small eta makes model perform very poorly despite relativly many epochs

---