David Steinmann

What’s in the Bottle? A Survey and Roadmap of Concept Bottleneck Models CBMs are a rapidly growing direction in interpretable-by-design machine learning. However, the field has become increasingly fragmented. Preprint Link: researchgate.net/publication/40…

#NeurIPS2025 has passed, but we hope the celebration will last forever. Here is a poster presentation we recorded at the event and we hope it can last forever in cyberspace. Paper: Object Centric Concept Bottlenecks Thanks for the authors: @Dav_Steinmann, @WolfStammer, @toniwuest, @kerstingAIML

Work together with @felixdivo @mkraus_io @toniwuest @LukasStruppek @felix_friedri @kerstingAIML

Curious about the differences between shortcuts, spurious correlations, and confounders? 🤔 Our new paper explains similarities and differences and introduces a comprehensive taxonomy of shortcut learning, reviewing the state of the field. Check it out! arxiv.org/abs/2412.05152

Our taxonomy includes: 🔍 Methods for shortcut detection 🛠️ Methods for shortcut mitigation 📊 A collection of datasets for research on shortcut learning If you think we’ve missed anything important, don't hesitate to reach out!

Excited to be at #NeurIPS2024 this week presenting our work Neural Concept Binder! 🤗 Stop by our poster to see how we derive expressive concept representations from unlabeled images. ⏰ Thu, Dec 12 11am–2pm 📍 East Hall A-C, #2103 See you there! 🎉✨

💡 Did you know that "Graph Neural Networks Need Cluster-Normalize-Activate Modules"? That's the title of the work we present at @NeurIPSConf this week! Our simple CNA modules added to each layer effectively mitigate oversmoothing in GNNs and strongly improve performance! 📈

Our recent workshop paper puts Vision-Language Models like #GPT4o to the test with Bongard problems🧩 We show that VLMs struggle to identify concepts that are quite intuitive to humans. Still a long way to go for human-like visual reasoning! 🤖🧠 arXiv: arxiv.org/abs/2410.19546

Very happy that our work Neural Concept Binder got accepted at @NeurIPSConf! We introduce a framework for interpretable and revisable concept learning in unsupervised scenarios 👉 arxiv.org/abs/2406.09949 Looking forward to presenting the work and exchanging ideas 🤗

📢 Presenting our latest research on #XAI and Interactive Learning at #ICML2024 🎉 Poster session: Tuesday from 11:30-1:00 (#2604). Title: Learning to Intervene on Concept Bottlenecks 🧠 paper: lnkd.in/eYgka8fj @Dav_Steinmann @felix_friedri @kerstingAIML

🚀 We present Neural Concept Binder for unsupervised symbolic concept discovery. It combines continuous and discrete encodings for concept representations that are: expressive ✔️ inspectable ✔️ revisable ✔️ 🤯 🔗arxiv.org/abs/2406.09949 @toniwuest @Dav_Steinmann @kerstingAIML

Our new XIT method pretrains 🏋️ on 75 time series 📈 datasets at once! 🔗: XD-MixUp interpolates between datasets 🧮: New SICC loss proposed 📊: Competitive on 5 baselines 🤝 Work with @mkraus_io, @Dav_Steinmann, @devendratweetin & @kerstingAIML! arxiv.org/abs/2402.15404

🚀 Excited to share our new work! Tackling time series confounders with RioT: 🔄 Feedback across time & frequency 🎯 Steers models to right reasons 📊 Tested on our novel production line dataset with natural confounding factors arxiv.org/abs/2402.12921 huggingface.co/datasets/AIML-…