6/
Speaking of which…
Our group at the Lamarr Institute focuses on Trustworthy AI for Healthcare. We're a bit of a unicorn: we love diving into the theory, working on things like generalization bounds and robustness through loss surface analysis,
(8/13)

Thinking about doing a PhD or Postdoc in ML?Don’t just look at Stanford, Oxford, or ETH.Look at Germany.
Yes, Germany.
(Thread 🧵)
(1/13)

After a long grey German winter finally a sunny evening by the Rhein - sunset, wine, and the gentle chaos of people learning to relax again. Bonn does it right.

At the ETIM '25 they had an artist produce graphical abstracts of all talks. I really love it! Especially the data-hungry model lurking in front of the data fridge... Thanks #IKIM for organizing this event.

7/ For all the nitty-gritty details, check out the full paper ("Little is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning"): arxiv.org/pdf/2310.05696 📖 #AI #MachineLearning #ResearchPublication

6/ What about interpretable models? They are often more trustworthy but often cannot be trained by federated learning. 🌲 FedCT doesn't discriminate against interpretable models. Works with decision trees, XGBoost, etc. Quality similar to centralized training. #InterpretableAI

5/💡Can we go even one step further? Federated Co-Training (FedCT) forms a consensus over local hard labels as a pseudo-labeling for the public dataset to augment local training data. 📊 FedCT has similar test accuracy as FL and DD with near-optimal privacy. #PrivacyProtection

4/ 🤝 Can we share less data while maintaining model performance?🌐In many fields, vast unlabeled datasets are publicly available. Think healthcare databases. 📈 We can leverage this to share information between clients: e.g., Distributed Distillation (DD) uses co-regularization.

3/ 🔒 Differential privacy offers a theoretical solution, but can it be practically applied? Our findings show that it improves privacy, but not too much, at the cost of model quality. Moreover, practical guarantees are poor (ε = 145, δ = 10−5, arxiv.org/pdf/2210.03843).

2/🏥 Data is often distributed and cannot be pooled - think healthcare. 🚫 Federated Learning may seem like the answer, but our practical results might surprise you: vulnerability against membership inference attacks is high (VUL is probability of successful attack). #HealthTech

1/ 🔬Our paper on Federated Co-Training got accepted at #AAAI2025. The approach leverages a public unlabeled dataset to share predictions of local models. It achieves a test accuracy similar to Federated Averaging with greatly improved privacy and can even train interpretable models. #trustworthyAI