Ebtisam A.

Bioinformatics Specialization · Cohort 2024

Distributed Dataloader for Spatial Transcriptomics

Spatial biology explores the organization and interaction of biomolecules like RNA, DNA, and proteins within tissues to gain insights into cancer, neuroscience, and disease. Spatial Transcriptomics, a key aspect of this field, focuses on how gene expression varies across tissue regions, aiding in Disease Mapping. This project aims to develop a Spatial Transcriptomics Dataloader to enhance model training scalability within a distributed ecosystem by generating mini-batches from tissue pucks using various sampling strategies. Conducted in a cell biology AI lab contributing to the Cellarium ML core library, the project involved utilizing distributed systems and single-cell analysis libraries, and implementing the dataloader under expert guidance. Key features include pre-computed KNN computations, chunk-based partitioning with exclusive data assignment, and sampling strategies like random and KNN subsampling to capture different length scales. The Dataloader supports multi-node and multi-GPU distributed model training and includes batch tracking for workflow monitoring.

Read the final report (PDF)

See it in the program report (PDF)

Artificial Intelligence Specialization · Cohort 2023

Solving Dialect Bias in Deep Neural Networks for Arabic Speech Tasks

In a world where language plays an influential role in shaping our identities, distinguishing dialects within a language is vital. Accurate dialect identification allows language technologies to adapt to regional nuances. However, existing models have biases and limitations, hindering effectiveness. Our aim is to enhance fairness, robustness, and generalizability, enabling models to accommodate diverse linguistic characteristics in Arabic dialects.

See it in the program report (PDF)