Marcell Winkler's Website

In this post, we explore a futuristic approach to secure, cloud-deployed medical image processing using Vision Transformers (ViTs). The paper BlindTuner outlines a framework that combines ViTs with homomorphic encryption to enable privacy-preserving fine-tuning — a powerful combination for clinical settings that require high accuracy and strict data confidentiality. [Read More]

This post discusses recent research exploring the over-smoothing effect in Vision Transformer (ViT) models — a phenomenon where deeper layers lose spatial detail and act like low-pass filters. A new study demonstrates how Fourier domain analysis and filtering can help retain critical high-frequency components, leading to more expressive and robust... [Read More]

This is a demo post to show you how to write blog posts with markdown. I strongly encourage you to take 5 minutes to learn how to write in markdown - it’ll teach you how to transform regular text into bold/italics/tables/etc.I also encourage you to look at the code that... [Read More]