Yahia Salaheldin Shaaban

Thesis collaboration: Developing a time series multimodal LLM

I was working under the supervision of MIT postdoc Dr. M. Umar B. Niazi (https://sites.google.com/view/umarniazi/home) leverging inference time adaptation (i.e low rank Hypernetwork as Meta-learning approach) to adapt learneable KKL obervers for non-autonomous systems.

- Developed a multi-spectral model for crop field segmentation. Built an internal multi-source satellite dataset benchmark, including Planet Labs and Landsat data. Supervised the annotation pipeline and implemented human-in-the-loop techniques to enhance model performance. - Optimized deep learning models using pruning and quantization. Deployed models on-premise and utilized Ray on top of Slurm, Docker, and TorchServe. - (Work in internship) Biomedical Imaging: Developed a data pipeline for DICOM images for a production-level segmentation model for breast cancer screening for Baheya hospital.

Implemented FedPara paper (FedPara: Low-Rank Hadamard Product for Communication-Efficient Federated Learning) during Flower Summer of Reproducibility.

- GNN-based model for predicting scientific breakthroughs: I have engaged in a GNN-based project that aims to predict scientific breakthroughs. By utilizing Graph Neural Network (GNN) approaches, I leverage the inherent structure of heterogeneous publication graphs while also incorporating the time dimension. This allows for the early-stage identification of significant scientific discoveries, contributing to the advancement of knowledge. Some aspects of my work include: 1- Extensive Literature Review: My research involves incorporating event modeling and prediction in fintech, trend prediction, and bibliometrics. This enables a comprehensive understanding of the emerging trends and key factors driving innovation in the financial technology sector. 2- Concept Extraction and Modeling: I employ various techniques such as topic modeling, semantic search, key phrase extraction, and concept knowledge graph construction. These methods help in organizing and extracting valuable insights from vast amounts of textual data, enhancing information retrieval and knowledge representation. 3- Large Scale Graph Modeling: Dealing with graphs of considerable size (~40 million nodes), I conduct centrality analysis, visualization, community detection, and construct citation, co-citation, and authorship graphs. Additionally, temporal graph analytics is employed to analyze the evolution of scientific networks over time. 4- Mining Large Scale Datasets: My research involves working with the Semantic Scholar Academic Graph Dataset and The Semantic Scholar Open Research Corpus. These large-scale datasets provide valuable resources for exploring and understanding patterns and relationships in academic research. - GNN in wireless technology: conducting research on using GNNs for enhancing certain aspects in 5G networks. Used NS3 framework to generate a dataset from lena dual stripe scenerio

Responsibilities: • Led Computer Vision subteam in 2022. • Developing the stabilization system of the ROV. • Developing Computer Vision Systems for underwater object detection and localization combining image processing and deep learning approaches that help ROV acts upon it autonomously. • Gave lectures on various software topics such as computer vision, machine learning, and developer tools. • Worked with Qt framework to develop a server-client architecture that connects ROV topside with AVR microcontroller. Achievements: • Team placed 1st in the MATE ROV Regional Competition in Egypt 2022. • Team placed 1st in Microsoft Azure Machine Learning ROV Challenge 2021. • Awarded Most Efficient Image Processing Award in the MATE ROV Arab Regional Competition 2021. • Team placed 1st in the MATE ROV Regional Competition in Egypt 2021. • Team placed 3rd in the MATE ROV International Competition telepresence and 1st in the underwater product demonstration. ------------ M.I.A ( Made In Alexandria ) is a robotics team founded in Alexandria, Egypt in 2011. For 10 years, the team has participated in major worldwide robotics competitions where it garnered multiple awards nationally and internationally.

Received extensive online training on various technologies: Data Warehouses, ETL, Big Data, Data visualization, Virtualization and Dockerization. Developed a hybrid sentiment analysis model built for business companies identifying clients impressions about products classifying data into positive and negative sentiments. ● Worked with multiple frameworks and libraries: flair, Spacy, PyTorch, Sklearn, Pandas. ● Analysed and profiled the performance of multiple models with multiple word embedding: fastText, Glove and transformers. ●Trained on various datasets including (Yelp, Amazon food and 140 sentiment). ● Utilized Roberta transformer. ● Applying transfer learning and cascading methodologies to enhance the model performance. ● Model sustained accuracy of 94%.

Working under the supervision of Dr. Salem Lahlou and Dr. Martin Takac. Mentored by Dr. Samuel Horvath.

Done one semester, took 3 courses: - Natural Language Processing (Same syllabus as Stanford CS224N. Course project: Arabic retrieval model with augmentation techniques, under the supervision of Dr. Nagwa ELmakky and Dr. Marwan Torki) - Reinforcement Learning (same syllabus as Berkeley CS285) - Simulation techniques

Bachelor thesis: Extended flower framework, leveraging Neural Collaborative Filtering to the personalized federated scheme using a bilevel optimization, and integrating Secure Multiparty Computation (Secure Aggregation underactive adversary scheme). Our model achieved better protection policies while enhancing model performance by 8% compared to the baseline.