Bashir Sadeghi
Machine Learning Researcher · PhD in Computer Science
Invariant & Fair Representation Learning · Theory & Experimental Evaluation · Scientific Machine Learning
I am a machine learning researcher with a PhD in Computer Science and Engineering from Michigan State University, awarded in May 2023. My work combines mathematical analysis, algorithm development, and controlled experiments to understand how learned representations preserve predictive information while limiting dependence on nuisance or sensitive attributes.
My PhD research focused on invariant and fair representation learning using kernel methods and neural networks. I studied utility–invariance and utility–fairness trade-offs, developed closed-form approaches to adversarial representation learning, and evaluated these methods against established baselines across datasets and experimental settings.
At the Facility for Rare Isotope Beams (FRIB), I applied self-supervised learning to radiation-detector pulse classification and parameter estimation. I worked with simulated and experimental measurements, evaluated performance under noise and overlapping signals, and investigated model failure cases. This work included Python/PyTorch and C++/LibTorch workflows for training, inference, and validation.
My research has appeared in CVPR, ICCV, TMLR, ECML-PKDD, IEEE Transactions on Signal Processing, and Nuclear Instruments and Methods A, with accompanying research software. I have also contributed to NSF-funded research and three U.S. patent filings arising from work at PARC.
Research Software & Projects
-
Invariant & Fair Representation Learning
Developed linear, kernel-based, and neural-network methods for learning predictive representations while reducing dependence on nuisance or sensitive attributes. Connected theoretical analysis of utility–invariance and utility–fairness trade-offs with baseline comparisons and controlled experiments.
Research implementations accompany work published in TMLR, ICCV, and ECML-PKDD, including closed-form approaches to adversarial representation learning.
Utility–Invariance Trade-offs Closed-Form Solvers Kernel Adversarial Learning -
SAEFit – Self-Supervised Learning for Radiation Detector Pulses
Developed a framework for pulse reconstruction, parameter estimation, and single/double-pulse classification at FRIB. Generated training examples from measured single-pulse traces through shifting, amplitude scaling, and superposition.
Evaluated against conventional signal-fitting methods on simulated and experimental measurements, with systematic analysis of noise, weak signals, temporal overlap, and failure cases. Implemented workflows using Python/PyTorch and C++/LibTorch.
Code & Data Paper -
Generalized Sampling & Reconstruction Tools
Implemented signal-reconstruction algorithms and kernel-based methods for periodically shift-variant systems, with experiment pipelines and companion software for published IEEE Transactions on Signal Processing and SampTA papers. -
Small Language Models on WikiText-2
Experimental LSTM and Transformer-style language models built with PyTorch Lightning, exploring tokenization, embeddings, causal attention, optimization, and training diagnostics.
An educational codebase with data pipelines, a small BPE vocabulary, model training, and evaluation on WikiText-2.
GitHub Repo -
Clinical Neuro Agentic RAG – Scientific Document Retrieval Prototype
Built a lightweight FastAPI prototype for document ingestion, chunking, embedding-based retrieval, and citation-aware answers to scientific or synthetic clinical questions.
Demonstrates retrieval-augmented generation and explicit handling of missing evidence. A research demonstration, not a clinical decision-making system.
GitHub Repo -
AI Audio Assistant
Built a browser and FastAPI workflow for transcription and summarization using speech and language-model APIs, including structured JSON responses and output validation.
GitHub Repo
Journal Publications
- On Characterizing the Trade-off in Invariant Representation Learning, TMLR 2022 Characterizes utility–invariance trade-offs and corresponding encoders in reproducing kernel Hilbert spaces. Featured Certification; Outstanding Certification Finalist.
- A Machine Learning Framework for Accurate and Robust Analysis of Radiation Detector Pulses, NIMA 2025 Self-supervised pulse reconstruction, parameter estimation, and multiplicity classification, evaluated on simulated and experimental detector signals.
- Constrained Sampling: Optimum Reconstruction in Subspace with Minimax Regret Constraint, IEEE TSP 2019
- Shifting Interpolation Kernel Toward Orthogonal Projection, IEEE TSP 2018
- Shift-variance and Nonstationarity of Linear Periodically Shift-Variant Systems and Applications to Generalized Sampling Processes, IEEE TSP 2016
Conference Publications
- Utility-Fairness Trade-Offs and How to Find Them, CVPR 2024 Studies attainable utility–fairness trade-offs and evaluates fair representation-learning methods.
- Adversarial Representation Learning With Closed-Form Solvers, ECML-PKDD 2021 Develops closed-form methods as an alternative to iterative adversarial optimization.
- Imparting Fairness to Pre-Trained Biased Representation, CVPRW 2020
- On the Global Optima of Kernelized Adversarial Representation Learning, ICCV 2019 Analyzes global optima and achievable utility and invariance in kernelized adversarial learning.
- Shift-Variance and Cyclostationarity of Linear Periodically Shift-Variant Systems, SampTA 2013
Patents
- Method and System for Optimizing a Pair of Affine Classifiers Based on a Diversity Metric
- Method and System for Creating an Ensemble of Machine Learning Models to Defend Against Adversarial Examples U.S. Patent 12,242,941 (2025)
- Method and System for Learning an Ensemble of Neural Network Kernel Classifiers Based on Partitions of the Training Data
Work Experience
- Research Associate, Michigan State University / FRIB (Sep 2023 – Mar 2026) Developed self-supervised methods for radiation-detector pulse classification and parameter estimation. Designed controlled evaluations using simulated and experimental measurements, analyzed failure cases, and built Python/PyTorch and C++/LibTorch workflows for preprocessing, training, inference, and validation.
- Research Assistant, Michigan State University (Sep 2017 – May 2023) Studied invariant and fair representation learning through mathematical analysis, algorithm development, and experiments. Developed linear, kernel-based, and neural-network methods, including closed-form adversarial learning approaches. Implemented baselines and reproducible comparisons, published research, and contributed to an NSF-funded project.
- Research Intern, Palo Alto Research Center (May – Sep 2020) Researched kernel-based ensemble models and implemented TensorFlow experiments comparing predictive accuracy, model diversity, and adversarial robustness. Contributed to three U.S. patent filings, including one issued patent.
- Deep Learning Instructor, Beaver-Edge AI (Jul 2026 – Present) Teach AI 410 online as a part-time contract instructor. Develop lectures and Python/PyTorch activities, assess programming assignments, and help students understand model behavior and debug implementations.
- Teaching Assistant, Michigan State University (2017 – 2021) Supported courses in programming, machine learning, computer networks, computer architecture, and deep learning.
Education
- PhD in Computer Science and Engineering, Michigan State University (May 2023) Thesis: Invariant Representation Learning via Functions in Reproducing Kernel Hilbert Spaces
- Master of Science in Electrical and Electronics Engineering, Eastern Mediterranean University
- Bachelor of Science in Electrical and Computer Engineering, Noshirvani University of Technology
Skills
- ML research: Invariant and fair representation learning, self-supervised learning, kernel methods, adversarial learning, optimization
- Experiments & evaluation: Experimental design, baseline comparisons, reproducible evaluation, quantitative error analysis, robustness and failure-case analysis
- Scientific data: Signal processing, time-series analysis, simulated and experimental measurements, noisy and overlapping signals
- Programming: Python, C/C++, MATLAB, JavaScript (basic)
- AI/ML: PyTorch, TensorFlow, LibTorch, NumPy, SciPy, Scikit-Learn
- Data: SQL, Pandas, data cleaning and visualization
- Software & Tools: Git, Linux, shell scripting, VS Code, Jupyter
- Web: FastAPI (REST APIs), HTML, basic CSS/JavaScript