Master’s Thesis: Model Fusion for Automated Data Annotation

Syntronic - A Global Design House; Kista; Sur place; SEK 167 000 - 234 000; Temps partiel

Résumé du poste

Syntronic - A Global Design House i Sverige söker en masterstudent inom ML/CS som vill utforska hur modellfusion och konsensusbaserade metoder kan förbättra automatiserad dataannotering.

Du kommer att arbeta med PyTorch och state-of-the-art modeller, bygga ett end-to-end annoteringsflöde och utvärdera noggrannhet, brus och tidsbesparing i jämförelse med manuell märkning.

Qualifications

• Kandidaten är en masterstudent inom ML/CS med stark Python- och DL-bakgrund.
• Erfarenhet av PyTorch/framväxande ML-ramverk är meriterande.
• Kunskap om bild- och textbaserade modeller ses som plus.

Responsabilités

• Implementera och utvärdera metoder för automatiserad dataannotering genom modellfusion och konsensusbaserade prediktioner.
• Bygga och benchmarka en automatisk annoteringspipeline som genererar högkvalitativa etiketter och flaggar tvetydiga fall för granskning.
• Analysera prestanda, brusreducering, osäkerhetskalibrering och tidsbesparing jämfört med manuell märkning.

Connaissances

Python

PyTorch

Deep learning

Computer vision

NLP

OpenCV

Hugging Face

Formation

Master student in ML/CS

Outils

SAM (Segment Anything)

Vision-Language Models

Description du poste

Do you want to explore the intersection of deep learning, model fusion, and automated data labeling?

In this thesis, we invite one or two students to develop and evaluate methods for automated dataset annotation using model fusion and consensus-based prediction techniques. The project combines predictions or representations from multiple pre‑trained foundation models (e.g., vision-language models, specialized detectors, or segmenters) to generate high‑confidence pseudo-labels, drastically reducing manual annotation effort while ensuring high dataset quality.

About The Project

The aim is to design and evaluate an automated annotation framework leveraging model fusion to generate reliable ground‑truth candidates. Within this scope, you will:

• Work with diverse domain datasets and state-of-the-art pre‑trained AI models.
• Implement and benchmark model fusion strategies (e.g., ensemble voting, multi-model consensus, weight merging, or confidence-weighted aggregation).
• Develop an automated annotation pipeline that generates high-quality labels and flags low-confidence/disagreed cases for human review.
• Analyze system performance in terms of annotation accuracy, noise reduction, uncertainty calibration, and time efficiency compared to manual labeling.

Possible Research Questions

• RQ1: How effectively can model fusion and multi-model consensus mechanisms improve the accuracy and reliability of automated annotations compared to single-model pseudo-labeling?
• RQ2: How can uncertainty estimation and disagreement metrics between fused models be leveraged to filter noisy labels and guide active human-in-the-loop verification?

Features & technologies

• Deep learning & computer vision / NLP (PyTorch, Hugging Face, OpenCV)
• Model fusion & ensembling techniques (Consensus Algorithms, Model Merging, Multi-agent/Multi-model pipelines)
• Automated annotation & Active Learning (Pseudo-labeling, Segment Anything/SAM, Vision-Language Models)
• Dataset quality evaluation & noise filtering

Who are you?

We are looking for one or two master’s students with an interest in machine learning, computer vision, and data engineering. Experience with Python and PyTorch/deep learning frameworks is valuable, but most importantly, you are curious, motivated, and eager to explore practical AI applications in a real-world context. A requirement for the position is the ability to communicate effectively in both Swedish and English, both verbally and in writing.

To give you the best possible support during your thesis, we’d like you to be able to come to the office connected to the project and spend most of your time working from there.