Open to Research & Applied Scientist roles

Katarzyna
Szymaniak

Machine learning for signals that won't hold still.

Choose your field — the page adapts
Tailored for wearables & health — the CV button follows. Switch any time.

I build biosignal and wearable models — sEMG, IMU, EEG — that still work on the next user, the next session and the next week. Not just on the benchmark. I build models for noisy, non-stationary, multichannel time series that hold up when the data drifts and labels are scarce — and the evaluations that prove it. Does an audio-language model actually listen — or lean on text priors and the device that made the recording? I'm taking my work on confounds and distribution shift to audio foundation models.

PhD · Edinburgh · 2026Ex-Meta Reality Labs (CTRL-labs)Collaborating with NVIDIA Research
See the research → CV · Health AI ↓ Email
01 — About

One problem, many signals

Most ML works when the world holds still. Real signals don't — so I build the models that keep up, and the evaluations that prove it.

I'm an ML researcher with a PhD from the University of Edinburgh (January 2026), supervised by Prof. Kianoush Nazarpour and Prof. Timothy Hospedales. My thesis — Mitigating Confounding Factors in Myoelectric Control Through Adaptive Modelling and Learning — sits where deep learning meets the inconvenient reality of biosignals: electrodes shift, postures change, and users adapt to the model while the model is trying to adapt to them.

At Meta Reality Labs (CTRL-labs) in New York, I took that problem to wrist-worn neural interfaces: fusing sEMG with IMU, and building few-shot personalisation for a pretrained Conformer decoder that had to hold up when the band came off and went back on.

My research targets the failure modes that decide whether a model survives deployment on continuous sensor data: non-stationarity, cross-user and cross-session shift, labels that cost human time, and signals corrupted by noise and motion. I've attacked them from three sides — a domain-generalisation model that needs zero data from the unseen condition, active learning that replaced a full calibration session with 3.2 minutes of queried labels, and benchmarks rigorous enough (~40k models, nested cross-validation) to separate real gains from noise.

I work end-to-end — human-subject protocol design, multi-session data collection (including amputee participants), modelling and statistical evaluation — and built data-processing and evaluation pipelines that were adopted across my lab. Since August 2026, in collaboration with NVIDIA Research, I've been asking my thesis question of audio foundation models: do they reason from the signal itself, or from text priors and the device and site that made the recording?

I'm looking for the team to do this with in earnest — biosignal or audio foundation models, health AI, neural interfaces, or any group that treats real-world robustness as the research problem, not the last ablation.

Same problem, different signal

The failure modes I work on, as they appear in each domain — and the evidence I have for each. The highlighted column follows the lens you picked above.

Problem
Wearables
Time series
Audio
What I've shown
The world moves under the model
Electrode shift, arm posture, sweat, fatigue
Regime change, sensor drift, broken seasonality
New room, noise floor, microphone
Quantified a ≤10% position-induced drop (GREAT); generalised to unseen positions with zero target data (DR-cGMM).
New user, new device
Cross-user, cross-session sEMG; re-donning
New sites, machines, patients
New speakers, accents, hardware
Few-shot per-user calibration of a Conformer decoder at Meta; cross-session learning curves over calibration budgets.
Labels are the bottleneck
User calibration time, clinic time
Expert annotation budgets
Transcription and preference labels
3.2 min of actively queried data beat a full calibration session (86% vs 82%).
More than one stream
sEMG + IMU at different rates
Multivariate sensors, asynchronous sampling
Audio + text
Early-fusion sEMG + IMU decoders; diagnosed modality imbalance — one stream dominating optimisation.
The model learns the wrong cue
Arm position leaking into gesture decoding
Site, device or sensor artefacts that correlate with the label
Recording device, site and text priors
Showed arm position is itself decodable from sEMG (~80%). Now testing whether health-audio LLMs rely on device and site rather than physiology.
Benchmarks flatter the model
Within-session splits
Look-ahead and cross-domain leakage
Test-speaker and test-device leakage
Leave-one-domain-out model selection, nested CV and Friedman/Nemenyi testing over ~40k models.
02 — Research

Three questions

Everything I work on comes back to three questions about deployed models on continuous signals. Each has peer-reviewed evidence behind it — and each transfers beyond the signal I first answered it on.

01 · Robustness

Distribution shift

How do you build a model that works in a condition it has never seen?

I treat recording conditions as domains and design for the unseen one: a class-conditional Gaussian mixture with a covariance regulariser shared across domains, which needs zero data from the target condition.

In your domain Arm position, band re-donning, long-term signal drift. Regime changes, new sites, slowly drifting sensors. Unseen microphones, rooms and noise conditions.
  • DR-cGMM — the only method to outperform ERM, MMD and CORAL
  • GREAT — an open benchmark that isolates one controlled shift
02 · Label efficiency

Adaptation on a budget

What is the smallest labelling budget that still moves the model?

Acquisition functions as a research object, not a utility: uncertainty-based querying (least-confidence, margin, entropy), few-shot personalisation, and parameter-efficient fine-tuning when labels cost human time.

In your domain Minutes of user calibration instead of full sessions. Spending expert annotation where it changes the model. Adapting to new speakers and devices from few labels.
03 · Rigour

Evaluation you can trust

Would this result survive deployment?

Splits that mirror deployment, model selection that never peeks at the test domain, and non-parametric significance tests — so that "method X helps" means something.

In your domain Cross-subject and cross-session splits by design. No look-ahead, no leakage across domains or entities. Held-out speakers, devices and acoustic conditions.
  • ~40k trained models, leave-one-domain-out + nested CV
  • Showed popular deep DG methods give no significant gain over ERM in the low-data regime
  • In progress: counterfactual tests for shortcut learning in audio-language models
03 — Now

Current work

Aug 2026 → present
Independent research · in collaboration with NVIDIA Research

Source shortcuts in health-audio language models Audio grounding and shortcut learning in multimodal LLMs Audio grounding and shortcut learning in multimodal LLMs

My thesis asked which confounding factors break EMG decoders. Health audio has its own: respiratory and auscultation recordings come from different devices and sites, and those differences can travel with the diagnosis. I'm investigating whether health-audio language models diagnose from physiology — or from the recording device and site. Multimodal LLMs can answer audio questions correctly for the wrong reasons. I'm investigating whether they reason from the audio signal or from text priors and spurious acoustic cues — the shortcut-learning face of the distribution-shift problem from my thesis. Multimodal LLMs can answer audio questions correctly for the wrong reasons. I'm investigating whether they reason from the audio signal or from text priors and spurious acoustic cues — the shortcut-learning face of the distribution-shift problem from my thesis.

Counterfactual evaluation
Change one factor of the input at a time and see what the model's answer actually depends on.
Encoder probing
Measure what frozen audio encoders represent about the recording source, not just the label.
Cross-dataset tests
Train on some sources, test on unseen ones: robustness under distribution shift, measured without leakage.
Counterfactual evaluation Audio-language models Shortcut learning

Planned next steps

Device-invariant pretraining

Self-supervised pretraining that pushes audio encoders to represent physiology rather than the recording device.

RL post-training for grounding

Post-training with GRPO and DPO, and SFT with LoRA, so the model is rewarded for answers that depend on the audio rather than on priors.

Agentic evaluation

Extending the evaluation from single answers to health-audio agentsagents that reason over audio across several steps.

The same question as my thesis: confounding factors, distribution shift, and evaluation that shows which cue the model is really using.

04 — Industry

Research at product scale

Meta Reality Labs · CTRL-labs · New York

Research Scientist Intern, Continuous Navigation

Sep 2022 – Mar 2023 · Host: Fabio Stefanini

The team's cross-user sEMG decoding research was published in Nature (2025) and underpins the Meta Neural Band.

  • Multimodal fusion. Extended the wrist-worn sEMG decoding stack to multimodal sEMG + IMU fusion for neuromotor AR/VR input: built the inertial data loading and preprocessing, trained early-fusion decoders, and diagnosed modality imbalance (one input dominating optimisation).
  • Few-shot personalisation. Built an end-to-end pipeline for new custom gestures — from the human-subject data-collection protocol to few-shot per-user calibration of a pretrained Conformer-based decoder (the convolution-augmented Transformer from speech recognition).
  • Evaluation under shift. Measured cross-session generalisation after the band is re-worn, with per-user learning curves over calibration budgets; presented results and limitations at the team's technical deep dive.
05 — Papers

Selected publications

Google Scholar →
EurIPS 2025Workshop · extended version under review at TMLR

Position: Domain Generalisation in Myoelectric Control

K. Szymaniak, K. Nazarpour, H. Gouk

Proposed DR-cGMM, a domain-regularised class-conditional Gaussian mixture model that generalises to unseen domains with zero data from them, via a covariance regulariser that shares structure across domains. It was the only method to outperform ERM, MMD and CORAL; popular deep domain-generalisation methods gave no significant gain over ERM in this low-data regime. Leakage-free benchmark: leave-one-domain-out model selection, nested cross-validation and Friedman/Nemenyi testing over ~40k trained models.

Domain generalisationBenchmarkingProbabilistic models
Scientific Data 2025Nature Portfolio · open dataset

EMG Dataset for Gesture Recognition with Arm Translation (GREAT)

I. Kyranou*, K. Szymaniak*, K. Nazarpour  * equal contribution

Designed the experiment and acquisition protocol for an open benchmark of 16-channel sEMG at 2 kHz with synchronised 18-sensor CyberGlove II hand kinematics, recorded over a 3×3 grid of 9 arm positions to isolate a controlled distribution shift. Quantified the shift: up to 10% accuracy drop across positions, and position itself decodable from the signal (~80%). Showed that position-aware hierarchical decoding fails outside an i.i.d. setting (position encoder 58%), motivating the domain-generalisation work above.

Open data · 21,000+ accessesControlled shiftsEMG + kinematics
Paper Data Code
Front. Neurorobotics 2022

Recalibration of Myoelectric Control with Active Learning

K. Szymaniak, A. Krasoulis, K. Nazarpour

First application of active learning to myoelectric control. 3.2 minutes of actively queried data beat a full calibration session (86% vs 82%; 12 able-bodied participants, ~2 points in 2 transradial amputees; offline simulation). Identified a failure mode: diversity-based batch selection degraded margin sampling, likely because informative boundary samples cluster and get discarded as redundant.

Active learningHuman-in-the-loop
Paper DOI
IEEE TNSRE 2026Survey

Deep Feature Learning from Electromyographic Signals for Gesture Recognition Systems

W. Zhong, X. Jiang, K. Szymaniak, M. Jabbari, C. Ma, K. Nazarpour

Survey of deep representation learning for EMG-based gesture recognition.

Paper DOI
Earlier research
  • Predicting behaviour from brain data (University of Edinburgh, 2021; Prof. Oisin Mac Aodha, with the Centre for Discovery Brain Sciences) — decoded the location of freely behaving mice from in vivo deep-brain calcium imaging, using class-activation-style attribution to identify the neuronal populations driving predictions.
  • Sequence models for EEG (Swansea University, 2020; Dr Jingjing Deng, Prof. Xianghua Xie) — deep sequential models for EEG in decision deadlocks, handling temporal dependencies, non-stationarity, cross-channel connectivity and noise, towards brain–computer interfaces.
  • HoloLens neuroimaging for medical education (Swansea University, 2019; Best Student Project award) — MRI brain scans rendered as 3D meshes and deployed on Microsoft HoloLens via Unity as a mixed-reality teaching tool. Demo ↗
06 — Background

The trajectory

Aug 2026 →

Machine Learning Researcher, independent research

In collaboration with NVIDIA Research · shortcut learning and grounding in audio-language models
2020 – 2026

PhD, Biomedical Artificial Intelligence

University of Edinburgh, School of Informatics · Sep 2020 – Jan 2026 · Supervisors: Prof. Kianoush Nazarpour, Prof. Timothy Hospedales · Thesis: Mitigating Confounding Factors in Myoelectric Control Through Adaptive Modelling and Learning
2022 – 2023

Research Scientist Intern, Continuous Navigation

Meta Reality Labs (CTRL-labs), New York · Sep 2022 – Mar 2023
2022

Oxford Machine Learning Summer School (OxML)

University of Oxford
2019 – 2020

MRes, Biomedical Artificial Intelligence

University of Edinburgh
2018 – 2020

MRes, Visual Computing

Swansea University
2015 – 2018

BSc (Hons) Computer Science, First Class Honours

Swansea University
07 — Writing

Notes & essays

Blog · Dec 2025

The Hidden Saboteurs: How Confounding Factors Keep Bionic Hands from Working in the Real World

A taxonomy of what breaks myoelectric control and why it's harder to fix than it looks.

Read
Blog · Aug 2025

Beyond the Cloud: SLMs, Local AI, and Agentic Constellations

A vision for local-first AI built on biology-inspired architectures.

Read
Notes · Ongoing

Academic Writing Tips and Resources

Habits and go-to resources for writing papers.

Read
08 — Contact

Let's talk

The next role.
A collaboration.
A good problem.

I'm looking for Research Scientist and Applied Scientist roles where models meet messy, real-world signals:

Wearables & health AI Neural interfaces Time-series & sensor ML Audio & multimodal LLMs Robustness & evaluation

Especially teams building biosignal or health-audio foundation models, personalisation, or validated on-device ML — where cross-user and cross-session shift is the product problem. Especially teams whose data drifts — sensors, operations, clinical records — and who care whether a model still works next quarter. Especially teams building or evaluating audio and multimodal foundation models, where grounding and robustness decide what ships. Based in Edinburgh, open to relocation.

Two CVs · same record, two emphases
Health AI & Wearables Biosignals, wearable sensing, health audio PDF · 2 pages Matches your field Machine Learning Distribution shift, time series, multimodal foundation models PDF · 2 pages Matches your field
LinkedIn Google Scholar