Modeling Non-Linear Data_Enhancing Accuracy with Multiple Approaches

Overview

Non-linear classification problems often require more sophisticated models than linear classifiers. This notebook explores how different neural network designs and hyperparameters (activation functions, learning rates, scheduling) can improve accuracy on a synthetic circular dataset.

Notebook Contents

1. Generate Synthetic Dataset
   Create a non-linearly separable dataset (concentric circles) using sklearn.datasets.make_circles.

2. Data Visualization
   Plot the dataset to illustrate the classification challenge.

3. Model Architectures
   Implement and train six distinct neural networks, varying:

   • Number of layers and units
   • Activation functions (ReLU, Sigmoid)
   • Learning rate schedules

4. Training & Evaluation
   Track loss curves, compute confusion matrices, and compare model accuracy.

5. Conclusion
   Summarize findings and best practices for non-linear classification.

---

Multi-Class Classification of Fashion MNIST

This repository contains a Jupyter Notebook titled Multi-Class Classification of Fashion MNIST.ipynb, which explores the Fashion MNIST dataset. The notebook includes data loading, visualization of class labels, and prepares the dataset for a multi-class image classification task.

Dataset Overview

Fashion MNIST is a dataset of 70,000 grayscale images of size 28x28 pixels, distributed as follows:

• 60,000 training images
• 10,000 test images
• 10 clothing categories, each labeled from 0 to 9:

Label	Class
0	T-shirt/top
1	Trouser
2	Pullover
3	Dress
4	Coat
5	Sandal
6	Shirt
7	Sneaker
8	Bag
9	Ankle boot

Source: Fashion MNIST – Zalando Research

Visualizations

The notebook includes the following visualizations to explore the dataset:

• Single Image Display
  View an individual training image along with its label (e.g., index 3 shows "Dress").

• Grid of Samples
  Display a 5x5 grid of 25 images from the training set with their corresponding class names to gain an intuitive understanding of the dataset.

Models Used

This project implements and compares multiple neural network models for Fashion MNIST classification:

• Model 1: Shallow Fully Connected Neural Network (Without Normalization)
  A basic MLP trained directly on raw pixel values.

• Model 2: Shallow Fully Connected Neural Network (With Normalization)
  Same as Model 1 but with input normalization to improve training dynamics.

• Model 3: Deep Fully Connected Neural Network (Learning Rate Finder)
  A deeper MLP architecture used to experiment with learning rate scheduling and optimization.

• Model 4: Tuned Deep Fully Connected Neural Network (Optimal Learning Rate)
  A fine-tuned version of Model 3 trained with the optimal learning rate for enhanced accuracy and convergence.

All models use fully connected (dense) layers. Convolutional architectures are not used in this version.

Evaluation

• Confusion Matrix Visualization
  Used to evaluate class-wise performance after model prediction on the test set.

• Random Image Prediction Plotting
  Displays a set of test images alongside their predicted and actual labels for intuitive model performance understanding.

---