Dr Simbarashe Aldrin Ngorima · Postdoctoral Researcher, MUST Deep Learning, NWU
I design efficient, interpretable neural networks for wireless channels, vehicles and the human gait. My doctoral work produced DPA-RDCNN, REACH and TurboQuant, methods that make deep models smaller, faster and easier to trust.
About
A vehicular channel and a walking human are both fast-changing time-series measured by imperfect sensors. I build the models that make sense of them.
I am a postdoctoral researcher in the MUST Deep Learning group at North-West University, affiliated with the Centre for Artificial Intelligence Research (CAIR) and NITheCS. My research develops gradient-based attribution methods and structured compression techniques for deep neural networks applied to time-series problems.
My PhD focused on channel estimation for IEEE 802.11p vehicular communication, producing five peer-reviewed publications and three IEEE journal submissions currently under review. The postdoctoral work carries the same methodology into deep learning for prosthetics, including gait phase prediction, metabolic cost estimation and inter-subject domain adaptation for wearable sensing.
Beyond research, I serve as Communications Chairperson for the Deep Learning Indaba, the largest machine learning organisation in Africa, and lecture Signal Theory to third-year engineering students.
Doctoral research
Vehicular channels change faster than any classical estimator can comfortably track, and the hardware they run on has no room to spare. These methods make deep channel estimators accurate, explainable and small — in that order.
vehicular channel estimation
A temporal convolutional architecture with data pilot-aided refinement for IEEE 802.11p channels. It tracks fast vehicular fading across six standard channel models and runs at 4.16 ms single-frame latency — fast enough for real deployment.
Read the paper →gradient-based interpretability
An interpretability framework that identifies which features and filters a channel estimator actually relies on, then turns that evidence into a principled guide for architecture compression across channel conditions.
Read the preprint →attribution-guided compression
Joint interpretability-guided pruning and near-optimal Lloyd-Max weight quantisation for deep vehicular channel estimators, shrinking models for edge deployment while preserving bit error rate performance.
Read the paper →Postdoctoral research
My postdoctoral work carries interpretability and compression methods from wireless communication into deep learning for prosthetics and exoskeletons.
Multi-source domain adaptation that lets gait models transfer across subjects, so a prosthetic controller trained on one group of people generalises to a new wearer.
Predicting the energetic cost of walking from wearable sensors such as EMG and joint kinematics, working towards assistive devices that adapt to how hard the wearer is actually working.
gait cycle · knee flexion · live phase readout
tap to pause
Teaching & service
Now
Publications
Contact · sync complete
I am open to research collaboration, speaking invitations and conversations about efficient and interpretable deep learning.