I'm a 24 years old Data Science Master student at EPFL
(École Polytechnique Fédérale de
Lausanne) in the field of Data Science, passionate about algorithmic and
Machine Learning. I’m ambitious and eager to learn, so I’m looking for an internship to put my acquired skills into application within
a dynamic environment.
I do a lot of sports, web 3.0 investing and I like to travel and discover new places on my free time.
Taha Zakariya
Chemin des Triaudes 7
1024 Ecublens, Switzerland
Phone number : +41 78 633 4603
Email : taha.zakariya@epfl.ch
In my current position as a Data Scientist Intern at Nestle R&D, I’m working on data processing
and classification techniques that will help identify environmental conditions that associate with multiple
phenomena of interest through advanced data analytics. My contributions have successfully improved previous predictions done by
a consulting company through the implementation of new processing techniques and predictive pipeline.
I also Developed a website to facilitate the data analysis of microbiome datasets and perform AI predictions for microbiologists with non-technical background.
Key words: Python, Statistical modeling, Feature extraction, Machine Learning.
Assisted 100+ students to understand theoretical concepts of analysis and physics given to first year Communication Systems engineers and Micro-engineers , providing them with the necessary resources to improve their skills.
A project which aims to promote connectivity of linear structures in 3D microscopy images using the idea of snakes active contour model by having a loss that minimizes the energy, instead of using a normal loss that tends to not take connectivity into account unless annotations are perfect.
An interdisciplinary project which goal is to participate in two international competitions: University Rover Challenge in USA and European Rover Challenge in Poland.
My task was to tackle a new aspect of the stereo odometry challenge with 360o cameras, using convolutional neural networks in order to determine the position of the rover.
The project aims to leverage the power of machine learning to replace the very expensive Lidar with normal cameras in the odometry estimation.
Minor in Management, Technologies and Entrepreneurship.
Main courses: Machine Learning, Applied Data Analysis, Statistics for data science, Deep learning, Systems for data science.
Very good academic results.
Selected for an exchange at Johns Hopkins University during the 3rd year (cancelled due to pandemic).
Main courses: Stochastic models, Probabilities and statistics, Object-oriented programming, Functional programming.
A preparatory year during which students focus on fundamental concepts of mathematics (analysis, linear algebra and analytical geometry), physics, computer science and chemistry.
Very good high school background, regular top 3 among all classes.
These 5 years at EPFL and the multiple hands-on projects I had the opportunity to work on allowed me to acquire many skills, some of which I present bellow.
Towards the end of my internship at Nestle, we decided to develop a website to facilitate
the data analysis of microbiome datasets. Specifically, the website is designed to compare two groups
of microbiome datasets (e.g., case vs control, positive vs negative etc) that are identified through the metadata.
It provides basic exploratory analysis and statistical tests tools. The website also provides various pre-processing
tools and Artificial Intelligence (AI) models, including SVM, Logistic Regression, KNN, and XGBoost, to help discover
potential patterns and insights in your data.
As in many areas of computer vision, deep networks now deliver state-of-the-art results for delineation tasks, such as finding axons and dendrites in 3D light microscopy images.
Most of the existing approaches rely on convolutional networks to extract from images binary masks denoting which voxels belong to neurites and which do not.
Unfortunately, they do not guarantee that the connectivity of the produced masks corresponds to that of the real neurite network.
The goal of this project is to experiment a new approach with recent brain data in order to promote connectivity of neurons in 3D predictions.
In the last five years, Bitcoin was by far the most volatile largely traded asset. Given its decentralized nature, multiple factors can manipulate its price direction. We wanted to shed light on one of those factors: Public Speakers' Quotations. On one side, retail investors stem their entry and exit strategies (when to buy and when to sell) from the feeling conveyed by the subjective speeches of notorious public investors, which may cause drastic fluctuations in the bitcoin's price and trading volume. On the other hand, news and people in social media are known to be very reactive to the cryptocurrency market fluctuations. Using the Quotebank dataset, we would like to investigate the correlation and the potential causality between the public speakers' quotes, the greed and fear index and Bitcoin’s historical data.
The Higgs boson is a particle whose existence is required to explain the standard model of physics. In order to detect it, physicists at CERN conducted particles collisions. However, the Higgs boson being rarely generated from these, many experiments were required and involved the creation of a huge amount of data. Moreover, when produced, the boson rapidly decay into different particles, making its detection difficult. Through this project, we aimed to classify the decay products of the collisions experiments between signal, corresponding to the presence of a Higgs boson, and background noise. To handle the complex analysis of the data, we used a machine learning approach, avoiding an analytical computation of the model.
Image registration is key in many medical areas, and especially in retinal images analysis as it deeply helps diagnostic from multimodal and multitemporal sources. This project explores some of the registration techniques applied to retinal images of a large private dataset: cross-correlation, feature-based and deep-learning algorithm. Vessel extraction is also key for some of these methods and take an important part of this study. Preliminary results of most techniques give a glimpse of what could be done for improvement, while we couldn’t obtain a satisfying deep-learning algorithm.
The goal of this project is to to implement a Noise2Noise model. A Noise2Noise model is an image denoising network trained without clean reference images, that learns to capture only the relevant in- formation of the noisy image, the denoised image. The state-of-the-art for this kind of problem is resolved by a UNet architecture. To perform this task, we train and optimize a model from noisy images. In this report we present our implementation consisting of a UNet with residual connections. We explain our approach and the different strategies that we used such as data augmentation and hyperparameter tuning of the model. In the end, we achieve a PSNR score of 24.6.
The goal of this project was to code a grid game where the character needs to solve many riddles in 3 game areas.
Try to win the game and follow the instructions of the README file to explore the different features !
The goal of this year's project, named Javass, is to write a program to play a variation of the famous
Jass game.
Jass is a very popular card game in Switzerland
and in other neighboring countries, and is close to belote and other similar games.
Processed a huge raw dataset by building ER and relational schemas, cleaning data (Python Pandas), loading it into DataGrip, realizing multiple complex queries (SQL) and optimizing them.
The project aims to leverage the power of machine learning to replace the very expensive Lidar with normal cameras in the odometry estimation.
Download Report
I did this task as a 2-day challenge.
I wrote an algorithm that is able to retrieve the corresponding context from all other contexts in the dataset based
on a question in the dataset. Several techniques were possible, I chose to go with information retrieval as a combination
between KdTree using TfIdf vectorization and sentence similarity using Hugging Face embeddings.