Deep Learning Techniques: Exploratory and Comparative Analysis

This repository contains a series of deep learning experiments and analyses conducted as part of a graduate-level course in Machine Learning. The focus of this work is to explore, implement, and optimize various deep learning models on structured and image datasets.

Repository Structure

File Name	Description
Data_Preprocessing_and_Exploratory_Analysis.ipynb	Preprocessing the dataset, handling missing values, and performing exploratory data analysis (EDA).
Neural_Network_Model_Training.ipynb	Designing and training a neural network for classification tasks, including performance evaluation.
Hyperparameter_Optimization.ipynb	Exploring the impact of different hyperparameters on model performance, such as dropout rate and optimizer choice.
CNN_Model_Comparison.ipynb	Comparing a custom CNN with modified VGG-13 architecture for image classification tasks.
Bonus_ResNet_Model.ipynb	Implementation and evaluation of a ResNet-based model for advanced classification tasks.
Model_Interpretability.ipynb	Techniques for interpreting and understanding neural network predictions.

Key Highlights

Data Preprocessing and EDA

• Cleaned the dataset by handling missing values and standardizing features.
• Performed visualization techniques like pair plots and correlation matrices to identify feature relationships.

Neural Network Implementation

• Designed a multi-layer neural network with ReLU activation, dropout, and batch normalization for regularization.
• Achieved a test accuracy of 79.82% using optimized hyperparameters.

Hyperparameter Tuning

• Experimented with dropout rates, batch sizes, and optimizers (Adam, RMSprop, SGD).
• Observed the best performance with a dropout rate of 0.5, batch size of 128, and RMSprop optimizer.

CNN and VGG-13 Comparison

• Implemented a custom CNN and modified VGG-13 architecture.
• VGG-13 outperformed the CNN model with a test accuracy of 92% compared to 89%.

Advanced Models and Techniques

• Implemented a ResNet-based model for enhanced feature extraction.
• Explored interpretability methods to analyze model predictions, including class-specific visualizations.

Performance Metrics

• Custom CNN: Test accuracy = 89%, Precision = 87%, Recall = 88%
• Modified VGG-13: Test accuracy = 92%, Precision = 91.96%, Recall = 91.62%, F1-Score = 91.57%
• ROC curves and confusion matrices were generated to validate the models’ performance.

Tools and Libraries

• Programming Language: Python
• Deep Learning Framework: PyTorch
• Other Libraries: NumPy, pandas, Matplotlib, Seaborn, scikit-learn