ML in PL

Generative models are powerful. The harder questions are whether we understand what's happening inside them, whether we can control it, and whether we can trust them with sensitive data. These three recordings tackle exactly that.

🎓 Kamil Deja — SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders
Most methods for removing unwanted concepts from diffusion models lack transparency. SAeUron uses sparse autoencoders trained on model activations to learn interpretable, concept-specific features — and intervenes precisely on those to block targeted content. It outperforms existing approaches on the UnlearnCanvas benchmark, handles multiple concepts with a single SAE, and holds up under adversarial attack.

🎓 Antoni Kowalczuk — Privacy Attacks on Image AutoRegressive Models
Image autoregressive models have quietly matched diffusion models on image quality while being faster to generate. The privacy risks, however, are significantly higher: Antoni's membership inference attack reaches 86.38% TPR at FPR=1%, compared to 6.38% for diffusion models. Dataset membership can be inferred from as few as 6 samples, and hundreds of training images can be directly extracted.

🎓 Łukasz Staniszewski — Controlling Generative Models through Parameter Localization
Less than 1% of a diffusion model's parameters govern its textual content in image generation. Building on an ICLR 2025 paper, Łukasz presents a unified framework for localizing and modulating these components across image, text, and audio models — enabling fine-grained editing, efficient fine-tuning, and control over musical attributes like tempo, instrumentation, and vocal style.

Links in the comments 👇

The quality of a research programme is shaped not only by speakers, but by the people in the room.
MLSS R&S 2026 is a five-day programme in Kraków focused on reliability and safety of machine learning methods and systems. The school brings together PhD students and researchers for invited lectures, discussions, and close interaction with speakers and participants.
As we approach the final week of registration, applications remain open until this Sunday (April 19, AoE).
Each edition is shaped not only by the programme, but by the group itself - often leading to in-depth discussions, new ideas, and ongoing collaborations.
If you know someone working in this area, this might be a good moment to share this with them.

📍 Kraków, Poland
📅 June 29 – July 3, 2026

Details in the first comment.