Neural Network from Scratch in MATLAB

Introduction

This repository provides a basic implementation of a neural network from scratch using MATLAB as a personal reference. It serves as supplementary material for CODECERDAS-6G project, Telkom University, Bandung, Indonesia.

How to use?

Directory Structure

├── func/
│   ├── activation/    % Activation functions
│   ├── backprop/      % Forward/backward propagation
│   ├── utils/         % Helper functions
│   └── weight_init/   % Weight initialization methods
├── xor_operation_train_no_bias.m
├── xor_operation_eval_no_bias.m
├── xor_operation_train.m  
└── xor_operation_eval.m

Configuration Guide

1. Layer Structure

   % Define layers (example: 2 inputs → 2 hidden → 1 output)
   input_size = N;       % N = number of inputs
   dense1 = H;           % H = number of hidden neurons
   dense2 = M;           % M = number of outputs

2. Matrix Dimensions

   % Weight matrices
   W1 = [H × N]          % Hidden layer weights
   W2 = [M × H]          % Output layer weights
   
   % Bias vectors
   WB1 = [H × 1]         % Hidden layer bias
   WB2 = [M × 1]         % Output layer bias
   
   % Data format
   INPUT = [samples × N] % Each row is one sample

3. General Rules

   • Hidden layer size: dense = any number
   • Weight matrix rows = number of neurons in current layer
   • Weight matrix cols = number of inputs to that layer
   • Bias vector length = number of neurons in that layer

4. Activation Functions

   activation: @func_activation
   derivative: @func_activation_deriv

5. Network Flow

   % forward pass pattern
   layer = func_forward_pass(input, W, b, WB, dense, @activation);
   
   % calculate error 
   next_e = target - layer;
   
   % backward pass pattern
   % for the last layer (output), set next_bias = 0
   [W, e] = func_backward_pass(layer, next_b, input, W, WB, b, next_e, l_rate, @activation_deriv);

Note: if using bias, ensure to add the bias input and weight.

Contact

For any inquiries or feedback, you can reach me at moefqy@rocketmail.com.