Machine Learning Projects

Selected machine learning and deep learning projects demonstrating experience in computer vision, generative AI, and time-series forecasting.

About These Projects

I am a PhD student in Electrical and Computer Engineering at The University of Alabama with research experience in machine learning, RF systems, and electromagnetics.
The projects below demonstrate practical implementations of deep learning models for computer vision, generative AI, and time-series prediction using Python, TensorFlow, and PyTorch.
These projects explore neural network design, transfer learning, multimodal models, and model interpretability techniques.

Machine Learning Skills

Programming: Python, NumPy, Pandas
Machine Learning: Scikit-learn, TensorFlow, PyTorch
Deep Learning: CNN, LSTM, GAN, Transfer Learning, Vision-Language Models
Computer Vision: Image classification, CNN architectures
Generative AI: Conditional GANs, Vision-Language Models
Model Evaluation: Cross-validation, learning curves, feature importance (SHAP)
Visualization: Matplotlib, training diagnostics

SELECTED Machine Learning Projects

Project 1

CNN Gender Classification from Facial Images

Problem

Develop a deep learning model capable of classifying gender from grayscale facial images.

Methodology

Reconstructed facial images from a pixel-based dataset and reshaped them into 48 × 48 grayscale images.
Designed a Convolutional Neural Network (CNN) architecture consisting of convolutional layers, pooling layers, and fully connected layers for hierarchical feature extraction.
Applied convolution and pooling operations to capture spatial patterns such as facial structure, edges, and textures.
Trained the model using TensorFlow/Keras, optimizing the network through backpropagation.
Monitored model performance using training and validation accuracy and loss curves to evaluate convergence and detect overfitting.

Results

The CNN model successfully learned discriminative facial features for gender classification and demonstrated stable training and validation performance.

Technologies

Python, TensorFlow, Keras, NumPy, Pandas, Matplotlib

Repository: GitHub

Dataset Samples (UTKFace Dataset): Examples of facial images reconstructed from the dataset used for CNN training.

CNN Architecture Overview

Project 2

CLIP + LoRA Image Classification (Vision-Language Model Fine-Tuning)

Problem

Pretrained multimodal models such as CLIP (Contrastive Language–Image Pretraining) are powerful but expensive to fully fine-tune.

The objective of this project is to explore parameter-efficient fine-tuning techniques that adapt large vision-language models for specific image classification tasks while minimizing the number of trainable parameters.

Methodology

Dataset Preparation: Used the Food-101 dataset, containing 101 food categories. Images were resized and preprocessed before being used for training and evaluation.
Pretrained Model: Utilized the CLIP vision-language model, which learns joint representations of images and text using contrastive learning.
Parameter-Efficient Fine-Tuning: Implemented LoRA (Low-Rank Adaptation) to adapt the pretrained CLIP model without updating all network parameters. This approach significantly reduces computational cost while maintaining strong performance.
Linear Probing: Applied linear probing on top of the frozen CLIP image encoder to evaluate baseline classification performance.
Model Training: Fine-tuned the model using PyTorch, updating only the LoRA adapter parameters while keeping the original CLIP weights frozen.

Results

The LoRA-based fine-tuning approach successfully adapted the pretrained CLIP model to the Food-101 classification task while training only a small subset of parameters.

This demonstrates how parameter-efficient fine-tuning techniques can effectively specialize large multimodal models for downstream tasks.

Technologies

Python, PyTorch, CLIP, LoRA, HuggingFace, NumPy, Matplotlib

Repository: GitHub

Project 3

Stock Price Forecasting using Machine Learning and LSTM

Problem

Financial markets produce large volumes of time-series data that can be analyzed to identify trends and patterns.

The objective of this project is to develop machine learning models capable of forecasting stock closing prices using historical market data.

Methodology

Data Collection: Collected historical stock market data containing features such as opening price, closing price, trading volume, and other market indicators.
Data Preprocessing: Cleaned and prepared the dataset by handling missing values, normalizing features using MinMax scaling, and organizing the data into chronological sequences suitable for time-series modeling.
Feature Importance Analysis: Applied SHAP (SHapley Additive exPlanations) to identify the most influential features contributing to stock price predictions.
Model Development: Implemented multiple machine learning models for comparison:

Linear Regression
Dense Neural Network
Long Short-Term Memory (LSTM) network

Model Training and Evaluation: Split the dataset chronologically into training, validation, and test sets.
Evaluated model performance using metrics such as Mean Absolute Error (MAE) and Mean Squared Error (MSE).

Results

The LSTM model demonstrated improved forecasting capability by capturing temporal dependencies in stock price movements.

Compared to traditional regression models, the LSTM network produced more accurate predictions for sequential financial data.

Technologies

Python, TensorFlow, Keras, Pandas, NumPy, Scikit-lear, SHAP, Matplotlib

Repository: GitHub

Explore More Projects on GitHub ↓

Page updated

Google Sites

Report abuse