Neural Network

This repository is a self-learning project on neural networks, covering both the foundations and practical implementations.
It combines theoretical understanding with applied experiments in fraud detection, computer vision, and sequence modeling.

Overview

The notebooks inside walk through key concepts in machine learning and deep learning, including:

• Neural networks and their structure
• Activation functions and gradient descent
• Multilayer architectures and backpropagation
• Overfitting and model evaluation
• Convolutional Neural Networks (CNNs) for image tasks
• Recurrent Neural Networks (RNNs) for sequence tasks

Basics

Biological Inspiration

• Human brain neurons fire when electrical signals cross a threshold.
• Artificial Neural Networks (ANNs) mimic this behavior.
• ANN = nodes (neurons) + weighted edges + biases + outputs.

---

Core Components

• Weights (W): Control feature importance.
• Bias (b): Adjusts activation threshold, allows flexibility.
• Hypothesis function: ŷ = f(Wx + b)
• Model's "guess" for outputs.
• Loss function: Measures difference between prediction (ŷ) and true target (y).

---

Gradient Descent

• Optimizes weights and bias by minimizing loss.
• Update rule:
  W := W - α * ∂Loss/∂W
b := b - α * ∂Loss/∂b

Types:

• Batch GD: Uses entire dataset.
• Stochastic GD (SGD): Uses 1 sample at a time.
• Mini-batch GD: Uses small random batches (most common).
• Backpropagation: Algorithm to calculate gradients efficiently across layers.

---

Activation Functions

• Step function: Outputs 0 or 1 (rarely used today).
• Sigmoid: Smooth curve in range (0, 1), used for probabilities.
• ReLU: 0 if x < 0, x if x ≥ 0, very common in modern networks.

---

Deep Neural Networks (DNNs)

• Multiple hidden layers stacked together.
• Each layer transforms inputs into higher-level representations.
• Final layer outputs probabilities for classification.

---

Regularization: Dropout

• Randomly disables some neurons during training.
• Prevents overfitting and improves generalization.

---

Applications

Fraud Detection using banknotes.csv

• Data loading, train-test split, model building, training, and evaluation

Computer Vision with edge detection

• Input image: edgedetection_input.input
• Processing, visualization, and saving output

Edge Detection output image:

edgedetection_output.png

• Handwriting Recognition with CNNs on MNIST
• Pre-trained model: MNIST_cnn_model (accuracy: 0.9887)
• Retrainable with custom naming (do not append .keras extension)

Trained MNIST output Model

MNIST_cnn_model.keras

---

Files

required files are in this repository

• requirement.txt
• banknotes.csv
• edgedetection_input.input
• python file neural_network.py
• NERVMAP.txt