Portfolio

---

Applications

Image Reconstruction form AutoEncoder

An application of autoencoders includes image denoising. I have used the persons dataset which is a subset of the COCO dataset. I have used the images present in the dataset and generated labels by adding salt and pepper noise to the image. The input to the model will be the noisy image created through using functions from opencv library and the output is the original image. In the process, the model learns to denoise the images. Thus, through a synthetic dataset, I have created an image denoiser.

AutoEncoder Image Processing Computer Vision Keras Deep Learning

—

Spam SMS Text Classification

Uses deep learning to detect spam messages. A Bi-Directional Encoder Representation from Transformers (BERT) as a feature extraction backbone is used to predict if the text of a given SMS is likely to be a Spam message.

BERT Data Cleaning Text Classification Keras Deep Learning

—

Modeling Political Conflicts

The objective of this project is to develop a predictive model to determine the likelihood of conflicts, civil unrest, or political violence in a country. By analyzing historical data and various socio-economic, political, and demographic factors, the model aims to provide early warning indicators and insights that can help policymakers and organizations proactively address potential conflicts.

Support Vector Machine (SVM) Deep Learning Random Forest Classifier Data Cleaning Feature Engineering Scikit Learn Tensorflow Keras

—

Semantic Search and Keyword Knowledge Graph

Shows underlying correlation patterns between various research fields and researchers in major Universities of Japan in the form a graph networks. The projects deal with multi lingual data including English and Japanese. It was built on top of pretrained BERT model from HuggingFace and Facebook’s FIASS indexing.

BERT NLP Graph Network Indexing Indexing Big Data FIASS Indexing Correlation



—

Geo-Spatial Analysis of Earthquakes in Turkey

Looks at the historic earthquake data, uses geospatial analysis to theorise the origin of earthquakes, and recommends most vulnerable districts in Turkey.

Data Analysis Geo-spatial Analysis Visualisation

---

Algorithm Implementations

Upper Confidence Bound Implementation Fom Scratch

Based on the assumption that given a set of possible options each option has an underlying normal distribution, Upper Confidence Bound Strategy is coded from scratch to choose the optimal option without knowing those distribution, with minimal exploration, and maximum exploitation. This algorithm is useful in a wide variety of problems such as marketing, drug efficacy, and recommender systems.

Reinforcement Learning Upper Bound Confidence Recommender System Machine Learning Algorithm

---

Naive Bayes Implementation From Scratch

From scratch implementation of Naive Bayes Classification Algorithm. Classifies a given set of datapoints into a certain number of classes.

Classification Machine Learning Algorithm

---

Logistic Regression From Scratch

From scratch implementation of Logistic Regression Classification Algorithm. Also, includes built it visualisation and validation functions.

Classification Machine Learning Algorithm Linear Model Regression

---