AHmed Mohammed

data scientist postdoctoral fellow at jefferson lab

United States

florida international university

Aug 2017 - Jul 2025 -7 yrs, 11 months

Job Details:Utilizing GNNs combined with functional connectivity (FC) analysis and attention mechanism to detect EEG epileptogenic spikes. A high average precision (AP) (above 0.97) was achieved compared to vanilla self-attention, hierarchical attention, and vanilla VGG models (below 0.92). Developing epileptic spike detector that is fed with features extracted from the standard four EEG frequency bands using MATLAB and PyTorch Lightning outperforming existing models (0.88 vs 0.8 ROC-AUC). Achieving high accuracy (above 0.95) in 3D semantic segmentation of MRI (magnetic resonance imaging) using standard 3D U-Net on Keras. Significantly enhancing the localization of epileptogenic sources upon investigating spectral power and FC patterns of epileptic EEG. Modeling a full-duplex decode-and-forward cooperative molecular communication system and building a MATLAB simulator to validate the developed model accuracy (above 0.95). Teaching: DSP applications and logic design labs.

istlink company

Jun 2016 - Jun 2017 -1 yr

Turkey

Job Details:As there was no certain scheduling algorithm adopted in LTE systems standard, there was a need to simulate, assess, and compare such algorithms. In my project, I built a MATLAB simulator to assess various QoS-aware scheduling algorithms. The goal was to identify an algorithm that efficiently assigns resource blocks to users in the OFDMA downlink of multi-carrier LTE systems in terms of throughput, packet loss rate (PLR), and fairness.

istanbul sehir university

Sep 2014 - Aug 2016 -1 yr, 11 months

Turkey

Job Details:Developing a software FM wireless channel model on MATLAB. Designing a hardware wireless channel emulator that avoids quantization noise effects by introducing time and frequency selectivity to the transmitted signal in the RF domain directly without the need for down/up conversions. Variable attenuators were used to output signals suffering the desired fading spectrum with the help of microcontrollers to imitate a wide range of fading channel types. Teaching: digital communication practice sessions and electric circuits lab.

jefferson lab

Job Details:The complicated nature of HPC systems and the associated large number of performance metrics makes it infeasible for system administrators to supervise. Hence, we adopt unsupervised techniques, i.e., (variational) autoencoders, to model the dynamics of the cluster running computationally intensive jobs. We investigated assisting such models with GNN blocks that allows exchanging information across the CPUs of a compute node. These models are compared in terms of their ability in capturing the salient features of the input as well as their ability to make distinction between the different jobs. The goal is to identify the best representative model that can be further adapted to the downstream task. We deploy such models in a human-in-the-loop production-based setting for the anomaly detection task where the champion model is identified from the daily competition held after the nightly training of the models. Formulating the problem of particle tracking in accelerators as a graph where the event hits on the detector layers are represented as nodes while track segments are represented as a subset of the graph edges that need to be correctly classified by the adopted GNN model. Significant speedups are achieved by batching multiple events which exploits the high parallel computational capability of GPUs.