Foundational models’ predictions (🦙♊️🦖) can propagate biases in causal downstream tasks, posing a significant risk in AI-supported scientific discovery (👩‍🔬🔎). Our solution: Causal Lifting of Neural Representation 🪜

Deeply inspired by @AIMS_Senegal1 students (and African colleagues) — not just for their genuine curiosity, but for their drive to apply ML to (less trendy) real social and environmental positive impact.

A true pleasure attending #MlssSenegal2025 in M’bour — a fun and intellectually vibrant gathering of world-class speakers and exceptionally driven students.

RT @TacoCohen: Nobody wants to hear it, but working on data is more impactful than working on methods or architectures.

RT @FrancescoLocat8: New preprint: "The third pillar of causal analysis? A measurement perspective on causal representations". https://t.co….

RT @pdebartols: Landed in Singapore for #ICLR—excited to see old & new friends! I’ll be presenting:. 📌 RAMEN @ Main Conference on Saturday….

RT @Dingling_Yao: Excited to present our paper “Unifying Causal Representation Learning with the Invariance Principle” at #ICLR2025 in Sin….

RT @unireps: 📢We are thrilled to introduce the ELLIS(@ELLISforEurope) x UniReps Speaker Series!📢. 🔵🔴 Each monthly session features a 40-min….

RT @riccardocadeii: @yudapearl If such model is perfect in prediction, yes. Otherwise, accuracy in prediction and in the causal downstream….

arXiv: Joint work with outstanding collaborators: peers @ilkerdemirel_ and @pdebartols, experimentalists Lukas Lindorfer and Sylvia Cremer, and my supervisors @CordeliaSchmid and @FrancescoLocat8. Illustration by Irene Guerrieri.

We hope this is a step for an alternative viewpoint to Causal Representation Learning: representations of data that make extracting causal information with some downstream estimator easier or at all possible.

This paper emphasizes the representation learning aspect of causal representation learning, ‘[. ] that make it easier to extract useful information when building classifiers or other predictors’ [Bengio et al. 2013].

For the first time, we zero-shot generalize causal inferences on ISTAnt dataset (without annotation) by causal lifting DINOv2 on our experiment variant.

We implement it via Deconfounded Empirical Risk Minimization (DERM), a new simple learning procedure minimizing the risk over a fictitious population, preventing potential confounding effects.

We propose suppressing the biases during fine-tuning on a similar reference experiment by blocking potentially spurious shortcuts.

We aim to support applied scientists in using deep learning to automatically annotate or make diagnoses on a target experiment to quantify the effect of an intervention, i.e., Prediction-Powered Causal Inference.

RT @pdebartols: Looking for a more efficient way to estimate treatment effects in your randomized experiment?. We introduce H-AIPW: a novel….

RT @Swarooprm7: @PavanJithesh PhD is once in a lifetime. Best time to be ambitious!!.

RT @sidgairo18: 🚨New Work Update🚨. "How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations" to be presented at u….

RT @ISTAustria: We are proud to announce that @FrancescoLocat8 has been selected for the #AI #Residency Program by @ChanZuckerberg. This pa….