Interval Adjoint Significance Analysis
A framework for estimating and bounding the significance of neural-network components using interval tangent and adjoint information.
- Interval arithmetic
- Adjoint methods
- Sensitivity analysis
Researcher & scientific software engineer
Sher Afghan
Malik
Scientific Computing · Automatic Differentiation · Machine Learning
About
Turning mathematical insight into working software.
My work connects mathematical ideas with practical scientific software: from sensitivity analysis and interval methods to GPU experiments and reproducible machine-learning pipelines. I am completing doctoral research in Computer Science at RWTH Aachen University under the supervision of Prof. Dr. Uwe Naumann.
I work across C++, Python, PyTorch, automatic differentiation, and HPC environments—implementing research algorithms, debugging numerical systems, and building reproducible GPU experiment workflows.
Professional profile →Research
Methods for understanding—and making—models efficient.
My doctoral research investigates how derivative information and rigorous interval bounds can reveal the significance of components inside neural networks.
Structured neural-network pruning
Using significance information to identify neurons or channels that can be removed while retaining useful model behavior.
- CNNs
- VGG
- ResNet
- Model efficiency
Gradient-aware retraining
Exploring bias compensation and Sobolev-style objectives that match model outputs and selected input derivatives after pruning.
- Higher-order derivatives
- PyTorch
- Fine-tuning
Projects
Research ideas, implemented.
Public work
Selected projects are being curated
Project descriptions will be added after the public repositories and their documentation have been reviewed. In the meantime, explore the source directly on GitHub.
Visit GitHub profile ↗Capabilities
A practical scientific computing toolkit.
From numerical kernels to multi-GPU experiments, I work across the full research-software loop.
Programming
- C++
- Python
- Bash
- LaTeX
Machine learning
- PyTorch
- TensorFlow / Keras
- CNNs
- VGG
- ResNet
- Model pruning
Numerical computing
- dco/c++
- Automatic differentiation
- Interval arithmetic
- Boost interval
- Higher-order derivatives
HPC & tooling
- Linux
- Slurm
- Multi-GPU workflows
- ONNX
- Git
- Experiment automation
Contact
Interested in research, engineering, or collaboration?
The best current way to connect is through my public GitHub profile. Additional professional contact links will be added once verified.
Find me on GitHub ↗