测试用户 的档案

# Applicant Background

> Comprehensive evidence-grounded profile synthesised exclusively from Stage 1 extracted facts.
> All claims are traceable to the source JSON. Gaps are explicitly marked as [Information not provided in source materials].
> Section 4 (Authoritative Skill Inventory) is a HARD CONSTRAINT for all downstream generation.

---

### SECTION 1: Summary Profile

Zhiyuan Chen (陈致远) is an M.Sc. candidate in Electrical Engineering and Information Technology (track: Signal Processing & Machine Learning) at the Technical University of Munich (TUM), Germany, with a final grade of 1.3 on the German scale. He completed his M.Sc. coursework in full and submitted his master's thesis, "Learned Compressed Sensing for Accelerated MRI: Initialization, Regularization and Uncertainty," in November 2025 (defense scheduled), supervised by Prof. Dr.-Ing. Ralf Berger at the TUM Chair of Signal Processing. He holds a B.Eng. in Information and Communication Engineering from Zhejiang University, China (GPA 3.9/4.0, ranked 12/268, top 5%).

Core methods and techniques: learned warm-start initialization for the FISTA ℓ1 solver with a fixed-point-aligned loss; compressed-sensing MRI reconstruction (ℓ1-regularized inverse problems, TV-regularized BPDN); unrolled optimization network baselines (MODL, 12 blocks); variational dropout / Bayesian per-pixel uncertainty maps; joint variational Bayesian inference with automatic relevance determination (ARD) priors; variational message passing; direction-of-arrival (DoA) estimation and sparse-array calibration (MUSIC, sparse recovery); software-defined-radio measurement campaigns (USRP B210, UHD); and deep micro-Doppler radar signature classification.

Research domains: accelerated MRI / medical imaging reconstruction; optimization-aware deep learning for inverse problems; uncertainty quantification for safety-critical sensing; array processing and DoA estimation under model mismatch; sensing and instrumentation with SDR/MRI hardware.

Publication status: two first-author preprints — "A Learned Initialization Scheme for Fast Compressed-Sensing MRI Reconstruction" (2025, Z. Chen, X. Lin, R. Berger; intended for IEEE signal-processing venues) and "Variational Bayesian Array Calibration under Sensor Model Mismatch" (manuscript under preparation, 2026, Z. Chen, R. Berger). Publication DOIs and impact factors are [Information not provided in source materials].

Languages: Mandarin Chinese (native), English (C1, IELTS 7.5), German (B1, B1.2 course in progress at the TUM Sprachenzentrum; willing to take German courses during the PhD).

Target role and countries: a funded PhD position (2026 intake) in signal processing, machine learning for inverse problems, sensing, or imaging systems, in Germany, the Netherlands, Sweden, Switzerland, Denmark, Belgium, or Finland.

**Explicit capability boundary (CANNOT credibly claim):** no wet-lab molecular biology, cell culture, PCR, cloning, sequencing, flow cytometry, or bioinformatics/computational biology; no MCMC expertise (his variational formulation is explicitly designed to avoid MCMC); no FPGA/ASIC/VLSI design, PCB design, or RF hardware fabrication; no CT/PET/cryo-EM reconstruction experience; no natural-language processing, large-language-model training, or reinforcement learning; no clinical-trial or regulatory-affairs experience; no professional-level German academic writing (German B1 only); Julia is at a basic level only; no CUDA/GPU-kernel development; no evidence of grant writing or independent funding acquisition.

---

### SECTION 2: Personal & Academic Profile

### Identity

| Field | Value |
|---|---|
| Full name | Zhiyuan Chen (陈致远) |
| Email | zhiyuan.chen.test@example.com |
| Phone | +49 000 000 0000 (test) |
| Mandarin Chinese | Native |
| English | C1 — IELTS 7.5 |
| German | B1 (B1.2 course in progress, TUM Sprachenzentrum); willing to take German courses during PhD |
| Target role | Funded PhD position, 2026 intake — signal processing, ML for inverse problems, sensing, imaging systems |
| Target countries | Germany, Netherlands, Sweden, Switzerland, Denmark, Belgium, Finland |

### Degree 1 — M.Sc., Electrical Engineering and Information Technology (Track: Signal Processing & Machine Learning)

- **Institution:** Technical University of Munich (TUM), Munich, Germany
- **Dates:** October 2023 – present (coursework complete; thesis submitted, defense scheduled); expected completion November 2025
- **Thesis title:** *Learned Compressed Sensing for Accelerated MRI: Initialization, Regularization and Uncertainty*
- **Supervisor:** Prof. Dr.-Ing. Ralf Berger (Chair of Signal Processing)
- **GPA:** 1.3 (German/Bavarian scale)
- **Status:** All coursework complete; master's thesis submitted

**Most relevant coursework (grade, ECTS):**
- Detection and Estimation Theory — A (1.0), 6 ECTS — *best of a cohort of 84*
- Statistical Signal Processing — A (1.0), 5 ECTS
- Compressed Sensing & Sparse Recovery — A (1.0), 5 ECTS
- Convex Optimization for Signal Processing — A (1.0), 5 ECTS
- Sensor Array Processing seminar — A (1.0), 3 ECTS
- Information Theory — A (1.3), 6 ECTS
- Machine Learning (IN2064) — A (1.3), 6 ECTS
- Statistical Learning for Big Data — A (1.3), 5 ECTS
- Deep Learning in Computer Vision — B (1.7), 8 ECTS
- Digital Communication — B (1.7), 6 ECTS
- Research internship: MRI reconstruction — passed, 6 ECTS

### Degree 2 — B.Eng., Information and Communication Engineering

- **Institution:** Zhejiang University, Hangzhou, China
- **Dates:** September 2019 – June 2023
- **GPA:** 3.9/4.0, ranked top 5% (12/268)

**Most relevant coursework:**
- Signals & Systems; Digital Signal Processing; Communication Principles
- Probabilistic Theory; Linear Algebra II; Algorithms & Data Structures
- Electromagnetic Fields & Waves; Microwave Technology; Optoelectronics
- Embedded Systems
- Radar Principles — project: micro-Doppler classification

### Target

- **Role type:** Funded PhD position (2026 intake)
- **Fields:** signal processing; machine learning for inverse problems; sensing; imaging systems
- **Countries:** Germany, Netherlands, Sweden, Switzerland, Denmark, Belgium, Finland

---

### SECTION 3: Research Projects

---

### Project 1 — Learned Compressed Sensing for Accelerated MRI: Initialization, Regularization and Uncertainty

**Institution:** Technical University of Munich (TUM), Munich, Germany
**Department/Group:** Chair of Signal Processing; compressed-sensing working group
**Title:** Learned Compressed Sensing for Accelerated MRI: Initialization, Regularization and Uncertainty
**Dates:** February 2025 – November 2025 (10 months; the supervisor records the whiteboard-to-manuscript cycle as approximately 8 months)
**Supervisor(s):** Prof. Dr.-Ing. Ralf Berger (TUM Chair of Signal Processing)
**Role:** Master's Thesis (author; first author of the resulting preprint)

**Situation/Background:**
- Magnetic resonance imaging (MRI) scan time is limited by the number of k-space samples that can be acquired.
- Compressed sensing (CS) reconstructs diagnostic images from undersampled data by promoting sparsity, but ℓ1 solvers require tens of iterations — too slow for clinical workflows.
- Existing fully learned unrolled networks cut iteration counts but require large labelled training cohorts and lose their connection to the underlying physical/optimization model.
- The project targets a middle path: retain the classical optimizer's guarantees while learning only what learning is good at (initialization and uncertainty).

**Task/Objective:**
- Develop a learned warm-start initialization for the classical FISTA solver, so that only 6 FISTA iterations (instead of 40) reach equal reconstruction quality.
- Train the initialization network with a loss aligned to the FISTA fixed point rather than to ground truth alone.
- Add a companion Bayesian variant with a variational-dropout head that produces per-pixel uncertainty maps flagging unreliable reconstructions.
- Evaluate on full fastMRI brain data, including out-of-distribution acceleration, and validate simulation-to-real transfer with a phantom experiment on a clinical 3T scanner.
- Own the whole chain independently: method conception, implementation, fastMRI evaluation, and organization of the 3T phantom campaign with the clinical partner.

**Actions:**
1. Conceived and designed a hybrid classical/learned pipeline in which a lightweight convolutional neural network maps zero-filled k-space data to a warm-start image that initializes the classical FISTA solver. Principle: any improvement in the classical solver directly benefits the learned scheme.
2. Implemented all training and evaluation pipelines in PyTorch; the warm-start network has 0.9 M parameters.
3. Trained on 3,500 fastMRI brain training slices with 500 validation slices at 8-fold undersampling.
4. Designed a fixed-point-aligned training loss (aligned to the FISTA fixed point, not to ground truth alone), which contributed +1.1 dB.
5. Added a data-consistency projection step, contributing +0.7 dB.
6. Benchmarked against a classical baseline: TV-regularized BPDN solved with 40-iteration FISTA.
7. Benchmarked against a learned baseline: MODL unrolled network with 12 blocks.
8. Ran an out-of-distribution generalization test at unseen 6-fold undersampling, comparing the degradation of the hybrid method against the unrolled baseline.
9. Re-derived the FISTA convergence argument himself when the convergence proof for the warm-start scheme stalled, and identified the missing strong-convexity assumption required to complete the argument.
10. Built a companion Bayesian variant using variational dropout in the warm-start network to produce per-pixel uncertainty maps.
11. Validated the uncertainty maps against true reconstruction error using Spearman rank correlation.
12. Organized and ran a phantom experiment on a clinical 3T scanner with the clinical partner to confirm simulation-to-real transfer of the learned initialization.
13. Wrote the manuscript and maintained the group's shared evaluation harness, on-boarding two younger students on it.
14. Discussed extensions to non-Cartesian sampling trajectories and parallel-coil (SENSE) models.

**Statistical approaches used:** SSIM, PSNR, Spearman rank correlation, FISTA convergence argument / fixed-point convergence proof with a strong-convexity assumption, uncertainty quantification.

**Results:**
- **SSIM 0.912** for the hybrid method vs **0.913** for full 40-iteration FISTA — statistically indistinguishable image quality.
- **PSNR 34.8 dB**.
- **1.2 s/slice** vs **4.9 s/slice** for full FISTA — a **4.1× lower wall-clock reconstruction time** at equal SSIM.
- **6 FISTA iterations** instead of **40** to reach equal reconstruction quality.
- **Out-of-distribution (6-fold undersampling):** SSIM drop of only **0.9 points** for the proposed method vs **4.2 points** for the unrolled (12-block MODL) baseline; the method **retains 97% of its in-distribution SSIM**, whereas the unrolled baseline loses most of its advantage.
- **Uncertainty validation:** per-pixel uncertainty maps correlate with true reconstruction error at **Spearman ρ = 0.71**.
- Ablations: fixed-point loss alignment **+1.1 dB**; data-consistency projection **+0.7 dB**.
- **Phantom experiment on a clinical 3T scanner** confirms simulation-to-real transfer of the learned initialization.
- **What was established:** that optimization-aware learning (learn the initialization, not the solver) gives clinically relevant speed-ups while preserving the classical solver's robustness and interpretability, and that a variational-dropout head can supply actionable per-pixel uncertainty for quality control.
- **Translational significance:** the thesis argues that optimization-aware learning is the practical route to **certifiable reconstruction software for clinical MRI**, because the classical solver remains in the loop and the uncertainty map supports the safety case.
- **Supervisor assessment:** Prof. Berger ranks him **among the top three master's students he has supervised in the last decade**; he took the learned-initialization idea from a whiteboard sketch to a submission-ready manuscript in **eight months**, including a full fastMRI evaluation and a 3T phantom campaign.

**Methodological Innovations & Troubleshooting:**
- Learned warm-start initialization for the classical FISTA solver — a hybrid design deliberately positioned between pure classical CS and fully unrolled learned networks.
- Fixed-point-aligned loss: the warm-start network is trained against the FISTA fixed point rather than ground truth alone, which is what allows a 6-iteration solver to match a 40-iteration solver.
- Variational-dropout head producing per-pixel uncertainty maps as a first-class output, not an afterthought.
- Independent mathematical troubleshooting: when the convergence proof for the warm-start scheme stalled, he re-derived the FISTA convergence argument himself and found the missing strong-convexity assumption.
- Designed the out-of-distribution evaluation protocol at 6-fold acceleration (different from the 8-fold training regime) to expose the fragility of unrolled baselines.
- Organized the 3T phantom campaign from scratch with the clinical partner — an experimental component outside the standard simulation-only thesis scope.

**Narrative Anchors:**
- **Medical imaging / accelerated MRI:** Owns the full stack from k-space sampling theory to a 3T phantom validation, so he can be handed an undersampled acquisition problem and deliver a clinically timed reconstruction with uncertainty.
- **Optimization-aware machine learning / inverse problems:** Trained a warm-start network against the FISTA fixed point and re-derived the solver's convergence proof, evidencing genuine mathematical depth rather than off-the-shelf deep-learning application.
- **Uncertainty quantification for safety-critical systems:** Produced and quantitatively validated (Spearman ρ = 0.71 against true error) a per-pixel uncertainty map, directly relevant to any position on trustworthy reconstruction.

---

### Project 2 — Variational Bayesian Array Calibration under Sensor Model Mismatch

**Institution:** [Institution not stated in source materials — associated with the TUM array-processing / sensor array processing context]
**Department/Group:** Array processing
**Title:** Variational Bayesian Array Calibration under Sensor Model Mismatch
**Dates:** [Information not provided in source materials]; manuscript under preparation with a stated 2026 year
**Supervisor(s):** Co-authored with Prof. Dr.-Ing. Ralf Berger (R. Berger); no formal supervisor role stated for this project
**Role:** First author (Z. Chen listed first)

**Situation/Background:**
- Sparse-array direction-of-arrival (DoA) estimation assumes calibrated sensor positions and gains.
- In practice, temperature drift and manufacturing tolerances introduce model mismatch that collapses super-resolution methods such as MUSIC and sparse recovery.
- Conventional calibration requires dedicated calibration sources, which is impractical in deployed arrays.

**Task/Objective:**
- Formulate array calibration as joint Bayesian inference over DoA angles and calibration parameters (gain, phase, position perturbation), solved by variational message passing.
- Develop a conjugate variational formulation that couples DoA and calibration variables **without MCMC**.
- Develop an automatic relevance determination (ARD) prior that detects dead or mislabeled elements.
- Provide an open evaluation protocol on both synthetic and SDR (software-defined radio) data.
- Own implementation, measurement campaign, and manuscript.

**Actions:**
1. Formulated the joint posterior over DoA angles and calibration parameters (per-element gain, phase, and position perturbation) and solved it by variational message passing.
2. Derived a conjugate variational formulation coupling the DoA and calibration variable blocks, explicitly avoiding MCMC to permit online operation.
3. Designed an automatic relevance determination prior over channel reliability; the ARD prior flagged 1 defective channel out of 16 in all trials.
4. Implemented the estimator in NumPy; per-iteration cost is **O(N³)** in array size, running 16 elements at **55 Hz on a laptop CPU**.
5. Built a synthetic evaluation on **16-element co-prime arrays** with **5% position jitter** and an SNR sweep from **−15 to 15 dB**, with sources at **1 kHz**.
6. Benchmarked against an uncalibrated MUSIC estimator as baseline (and against sparse-recovery super-resolution methods, which are also affected by model mismatch).
7. Implemented the SDR measurement campaign using **2× USRP B210** software-defined radios with the **UHD driver**, **8-channel synchronized**.
8. Established a ground-truth reference using a **mechanical rotator** with a **roto-tilter reference**, enabling over-the-air DoA accuracy assessment.
9. Quantified posterior uncertainty for each estimated angle as a native output of the variational formulation.
10. Wrote the manuscript and assembled the open evaluation protocol covering both synthetic and over-the-air data.

**Statistical approaches used:** joint Bayesian inference; variational message passing; conjugate variational formulation; automatic relevance determination prior; posterior uncertainty quantification; RMSE; RMS error against a mechanical reference; MCMC explicitly **not** used.

**Results:**
- **Synthetic:** on 16-element co-prime arrays with **5% position jitter**, the calibrated estimator keeps **RMSE below 1.2° down to −8 dB SNR**, where an **uncalibrated MUSIC estimator exceeds 10°** error.
- **Over-the-air:** with two USRP B210 software-defined radios, DoA tracks are recovered within **2.1° RMS** of the reference mechanical rotator.
- **Computational cost:** per-iteration cost **O(N³)** in array size; **16 elements run at 55 Hz in NumPy on a laptop CPU** — i.e. real-time-capable at this array size, without GPU.
- **Fault detection:** the ARD prior **flagged 1 defective channel out of 16 in all trials**.
- **Setup:** synthetic — co-prime array (16 elements), sources at 1 kHz, SNR sweep −15…15 dB; real — 2× USRP B210, 8-channel synchronized, roto-tilter reference.
- **What was established:** that robust super-resolution DoA estimation is achievable under realistic sensor model mismatch without calibration sources and without MCMC, while simultaneously quantifying posterior uncertainty and detecting faulty elements.
- **Translational significance:** online, source-free array calibration for deployed arrays — relevant to radar, wireless sensing, and any fielded sensor array subject to drift.

**Methodological Innovations & Troubleshooting:**
- No calibration sources needed — the calibration parameters are inferred jointly with the angles.
- Runs online (no MCMC), enabled by the conjugate variational construction.
- Quantifies posterior uncertainty for each estimated angle.
- Couples DoA and calibration variables without MCMC — the central algorithmic innovation.
- ARD prior detecting dead/mislabeled elements as an automatic diagnostic.
- Built the SDR measurement campaign from scratch, including the mechanical-rotator reference setup for over-the-air ground truth.
- Open evaluation protocol on synthetic + SDR data, supporting reproducibility.

**Narrative Anchors:**
- **Array processing / radar / DoA estimation:** Combines variational Bayesian inference with a hands-on USRP B210 measurement campaign, so he can both derive the estimator and prove it against a mechanical ground-truth reference.
- **Bayesian machine learning / variational inference:** Built a conjugate variational formulation that avoids MCMC and remains real-time (55 Hz at 16 elements on a laptop CPU) — relevant to any lab wanting scalable approximate inference in signal processing.
- **Robust sensing / model-mismatch tolerance:** Demonstrates a 10× reduction in low-SNR DoA error versus uncalibrated MUSIC, a directly transferable result for any array subject to temperature drift or manufacturing tolerance.

---

### Project 3 — Radar Micro-Doppler Signature Classification with Deep Networks

**Institution:** Zhejiang University, Hangzhou, China
**Department/Group:** [Department not stated in source materials]; conducted as a project within the Radar Principles course track
**Title:** Radar micro-Doppler signature classification with deep networks
**Dates:** March 2022 – June 2023 (approximately 16 months)
**Supervisor(s):** [Information not provided in source materials]
**Role:** Undergraduate research assistant

**Situation/Background:**
- Micro-Doppler signatures encode target kinematics (e.g. rotating parts, gait) and are used for radar target classification.
- Building a labelled radar signature dataset requires costly measurements, motivating synthetic data generation.
- The project addresses multi-class micro-Doppler classification with deep networks.

**Task/Objective:**
- Build a synthetic dataset generator to produce micro-Doppler training data.
- Train deep networks to classify micro-Doppler signatures across 8 classes.
- Own the dataset generation, model training, and evaluation end to end.

**Actions:**
1. Built a synthetic dataset generator in **MATLAB and Python** producing labelled micro-Doppler signatures for 8 target classes.
2. Trained **deep networks** on the generated signatures for multi-class classification.
3. Evaluated classification performance using the F1 metric, within the Radar Principles course project framework (micro-Doppler classification).

**Statistical approaches used:** F1 score; per-class evaluation across 8 classes.

**Results:**
- Achieved **94% F1 on 8 classes**.
- Established a working synthetic-to-model pipeline for radar micro-Doppler classification ahead of the master's degree.
- Delivered over a 16-month period (March 2022 – June 2023) alongside B.Eng. coursework, demonstrating sustained independent research involvement as an undergraduate.

**Methodological Innovations & Troubleshooting:**
- Built the synthetic dataset generator from scratch in MATLAB/Python, removing the dependency on scarce labelled field measurements.
- Extended a B.Eng. Radar Principles project into a full 8-class deep-learning classification study.

**Narrative Anchors:**
- **Radar / RF sensing:** Early, hands-on evidence of turning radar signal models into a labelled learning problem — directly complementary to his later SDR and array-processing work.
- **Deep learning for signal classification:** 94% F1 across 8 classes shows he can build training pipelines when data does not exist and must be simulated.
- **Undergraduate research persistence:** A 16-month project run alongside full-time coursework demonstrates the finishing behaviour noted later by his master's supervisor.

---

### Project 4 — MRI Reconstruction Research Internship

**Institution:** TU Munich, Munich, Germany
**Department/Group:** [Information not provided in source materials — MRI reconstruction]
**Title:** MRI reconstruction
**Dates:** [Dates not provided in source materials]
**Supervisor(s):** [Information not provided in source materials]
**Role:** Research internship (6 ECTS, passed)

**Situation/Background:**
- [Information not provided in source materials].

**Task/Objective:**
- MRI reconstruction research internship, completed within the M.Sc. programme.

**Actions:**
1. MRI reconstruction research work at TU Munich, credited with 6 ECTS and passed.

**Results:**
- Passed, 6 ECTS.
- Served as the entry point into MRI reconstruction research that led to the master's thesis under Prof. Berger.

**Methodological Innovations & Troubleshooting:**
- [Information not provided in source materials].

**Narrative Anchors:**
- **Medical imaging:** Formal early exposure to MRI reconstruction inside the same institution and field as his thesis.
- **Academic progression:** Provides a documented internship-to-thesis trajectory within one research line.

---

### SECTION 4: Technical Skills & Authoritative Inventory

### Depth-Expanded Skills

**Programming & Scientific Computing**
- **Python (PyTorch, NumPy, SciPy)** — daily driver; all master's-thesis training and evaluation pipelines implemented in PyTorch; the variational Bayesian array-calibration estimator implemented in NumPy (55 Hz, 16-element array, laptop CPU).
- **PyTorch** — implemented all training and evaluation pipelines; trained a 0.9 M-parameter warm-start convolutional network on 3,500 fastMRI slices.
- **PyTorch training at multi-GPU scale / distributed training** — CV-listed capability for scaled training; SLURM clusters listed for cluster job submission.
- **C++17** — CV-listed programming skill (no project-level detail provided).
- **MATLAB** — used to build the synthetic micro-Doppler dataset generator (MATLAB/Python) during the Zhejiang University radar project.
- **Julia** — basic level only.
- **Git, Linux, SLURM clusters** — CV-listed tooling for version control, Linux environments, and HPC job scheduling.
- **NumPy** — estimator implementation for array calibration; 16-element array at 55 Hz on a laptop CPU.

**Signal Processing & Estimation**
- **Compressed sensing** — ℓ1-regularized inverse problem formulation, sparsity-promoting reconstruction from undersampled k-space; coursework "Compressed Sensing & Sparse Recovery" (A/1.0, 5 ECTS).
- **Compressed-sensing MRI reconstruction** — fastMRI brain data at 8-fold undersampling; out-of-distribution test at 6-fold undersampling.
- **FISTA optimization / FISTA** — classical FISTA solver, 40 iterations as reference, reduced to 6 iterations with a learned warm start; re-derived the FISTA convergence argument and supplied the missing strong-convexity assumption.
- **Array processing / beamforming** — co-prime array geometry (16 elements), DoA estimation, MUSIC and sparse-recovery baselines; Sensor Array Processing seminar (A/1.0, 3 ECTS).
- **Detection & estimation theory** — A (1.0), 6 ECTS, best of a cohort of 84, taught by Prof. Dr.-Ing. Ralf Berger.
- **Statistical signal processing** — A (1.0), 5 ECTS.
- **Sampling theory** and **information theory** — Information Theory A (1.3), 6 ECTS.
- **Direction-of-arrival (DoA) estimation** — sparse-array DoA under model mismatch; sub-degree RMSE down to low SNR in simulation; degree-level RMS agreement against a mechanical-rotator reference.
- **Array calibration** — joint inference of per-element gain, phase, and position perturbation; robustness to 5% position jitter.

**Optimization & Inverse Problems**
- **Convex Optimization for Signal Processing** — A (1.0), 5 ECTS.
- **Learned optimization / hybrid classical-learned pipeline design** — learned warm-start initialization for FISTA, with the design principle that classical-solver improvements propagate to the learned scheme.
- **Fixed-point-aligned loss** — loss aligned to the FISTA fixed point rather than to ground truth alone; contributed +1.1 dB.
- **Unrolled optimization networks** — working command of the MODL unrolled baseline (12 blocks) as a benchmark and a critical understanding of its limitations (large labelled cohorts, weakened physical-model connection, larger OOD SSIM degradation).

**Machine Learning & Bayesian Methods**
- **Convolutional neural networks** — lightweight warm-start CNN mapping zero-filled k-space to an image; 0.9 M parameters.
- **Deep learning for inverse problems** — end-to-end learned reconstruction for undersampled MRI.
- **Bayesian uncertainty estimation / uncertainty quantification** — variational dropout head producing per-pixel uncertainty maps; validated against true reconstruction error by rank correlation.
- **Variational dropout** — implemented in the warm-start network to yield the companion Bayesian variant.
- **Variational Bayesian inference** — joint posterior over DoA angles and calibration parameters.
- **Variational message passing** — solution method for the joint Bayesian calibration model.
- **Conjugate variational formulation** — couples DoA and calibration variables without MCMC.
- **Automatic relevance determination (ARD)** — prior that flags dead/mislabeled elements; detected 1 defective channel out of 16 in all trials.
- **Statistical Learning for Big Data** — A (1.3), 5 ECTS; **Machine Learning (IN2064)** — A (1.3), 6 ECTS; **Deep Learning in Computer Vision** — B (1.7), 8 ECTS.

**Hardware, Instrumentation & Experimentation**
- **USRP B210 software-defined radio** — 2× USRP B210, 8-channel synchronized over-the-air measurement campaign.
- **UHD driver** — instrument control for the USRP B210 measurement campaign.
- **Software-defined radio (SDR) measurement** — implemented the full measurement campaign including the mechanical-rotator / roto-tilter reference for ground-truth DoA.
- **MRI phantom experiments** — ran the phantom experiment on a clinical 3T scanner, organized with the clinical partner, confirming simulation-to-real transfer.
- **Clinical 3T scanner** — phantom validation of the learned initialization.
- **Estimator implementation** — from derivation to NumPy implementation and measurement validation.

**Communication, Collaboration & Research Practice**
- **Scientific manuscript writing / manuscript writing** — wrote the first-author preprints, including the whiteboard-to-submission-ready manuscript in eight months.
- **Collaboration / mentoring** — maintained the group's shared evaluation harness and on-boarded two younger students on it.
- **Finishing / execution** — described by his supervisor as someone who "finishes," evidenced by the 8-month thesis cycle including a 3T phantom campaign.

**Domain Knowledge**
- **MRI reconstruction**, **fastMRI**, **non-Cartesian sampling** (discussed as a thesis extension), **parallel-coil (SENSE) models** (discussed as a thesis extension).
- **TV-regularized BPDN** and **MODL unrolled network (12 blocks)** as baselines.
- **Radar micro-Doppler classification** — 8-class deep-network classifier; synthetic d