shreyadavan/open-elective-fine-tuned-model

Open Elective - Custom Neural Network Fine-Tuning

Project Description

This project demonstrates training a custom neural network on a tabular student dataset followed by fine-tuning the same model to improve classification accuracy.

Model Architecture

The neural network contains:

• Input layer: 10 features
• Dense layer: 32 neurons, ReLU activation
• Dense layer: 16 neurons, ReLU activation
• Output layer: 1 neuron, Sigmoid activation

Training Process

Original Model

The original model was trained using:

• Optimizer: Adam
• Learning rate: 0.001
• Epochs: 3
• Batch size: 32
• Loss: Binary Crossentropy

Original Accuracy:

73.5%

Fine-Tuned Model

The same model was subsequently fine-tuned using:

• Optimizer: Adam
• Learning rate: 0.0001
• Epochs: 20
• Batch size: 32
• Loss: Binary Crossentropy

Fine-Tuned Accuracy:

80.0%

Accuracy Improvement

Model	Accuracy
Original Model	73.5%
Fine-Tuned Model	80.0%
Improvement	6.5 percentage points

Files

• fine_tuned_model.keras - Fine-tuned TensorFlow/Keras model
• scaler.pkl - StandardScaler used for preprocessing
• dataset.csv - Dataset used for training and evaluation
• Open_Elective.ipynb - Complete training and fine-tuning notebook

Method

The project follows:

Dataset → Preprocessing → Custom Neural Network → Original Training → Fine-Tuning → Accuracy Comparison

Result

The fine-tuned model achieved 80.0% test accuracy compared with 73.5% for the original model, giving an improvement of 6.5 percentage points.