Boston Housing Price Prediction — Epoch vs Loss

This project demonstrates how increasing the number of training epochs in a neural network reduces the training loss using the Boston Housing dataset.

Objective

To visualize and prove that training longer (more epochs) helps the model minimize loss — up to a certain point — before overfitting begins.

Dataset

The project uses the Boston Housing dataset, which contains information about housing prices in Boston suburbs.

• Features: 13 numerical attributes (e.g., RM, LSTAT, PTRATIO)
• Target: Median value of owner-occupied homes (MEDV)

Workflow Overview

1. Data Loading and Exploration
   • Load dataset using scikit-learn
   • Perform feature scaling
2. Model Definition
   • Simple feedforward neural network using PyTorch
3. Training
   • Train model with different numbers of epochs
   • Track loss across epochs
4. Visualization
   • Plot the relationship between Epochs and Loss

Key Results

• As epochs increase, training loss consistently decreases
• After a certain point, further epochs yield diminishing returns
• Demonstrates the concept of convergence in gradient descent

Visualization

The plot below shows how the loss decreases as the number of epochs increases:

Requirements

Install dependencies:

pip install torch torchvision matplotlib scikit-learn