We propose a novel formalism for describing Structural Causal Models (SCMs) as fixed-point problems on causally ordered variables, eliminating the need for Directed Acyclic Graphs (DAGs), and establish the weakest known conditions for their unique recovery given the topological ordering (TO). Based on this, we design a two-stage causal generative model that first infers in a zero-shot manner a valid TO from observations, and then learns the generative SCM on the ordered variables. To infer TOs, we propose to amortize the learning of TOs on synthetically generated datasets by sequentially predicting the leaves of graphs seen during training. To learn SCMs, we design a transformer-based architecture that exploits a new attention mechanism enabling the modeling of causal structures, and show that this parameterization is consistent with our formalism. Finally, we conduct an extensive evaluation of each method individually, and show that when combined, our model outperforms various baselines on generated out-of-distribution problems.

Foundation models have brought changes to the landscape of machine learning, demonstrating sparks of human-level intelligence across a diverse array of tasks. However, a gap persists in complex tasks such as causal inference, primarily due to challenges associated with intricate reasoning steps and high numerical precision requirements. In this work, we take a first step towards building causally-aware foundation models for treatment effect estimations. We propose a novel, theoretically justified method called Causal Inference with Attention (CInA), which utilizes multiple unlabeled datasets to perform self-supervised causal learning, and subsequently enables zero-shot causal inference on unseen tasks with new data. This is based on our theoretical results that demonstrate the primal-dual connection between optimal covariate balancing and self-attention, facilitating zero-shot causal inference through the final layer of a trained transformer-type architecture. We demonstrate empirically that CInA effectively generalizes to out-of-distribution datasets and various real-world datasets, matching or even surpassing traditional per-dataset methodologies. These results provide compelling evidence that our method has the potential to serve as a stepping stone for the development of causal foundation models.

Abstract This paper presents a digital image dataset of historical handwritten birth records stored in the archives of several parishes across Sweden, together with the corresponding metadata that supports the evaluation of document analysis algorithms’ performance. The dataset is called SHIBR (the Swedish Historical Birth Records). The contribution of this paper is twofold. First, we believe it is the first and the largest Swedish dataset of its kind provided as open access (15,000 high-resolution colour images of the era between 1800 and 1840). We also perform some data mining of the dataset to uncover some statistics and facts that might be of interest and use to genealogists. Second, we provide a comprehensive survey of contemporary datasets in the field that are open to the public along with a compact review of word spotting techniques. The word transcription file contains 17 columns of information pertaining to each image (e.g., child’s first name, birth date, date of baptism, father's first/last name, mother’s first/last name, death records, town, job title of the father/mother, etc.). Moreover, we evaluate some deep learning models, pre-trained on two other renowned datasets, for word spotting in SHIBR. However, our dataset proved challenging due to the unique handwriting style. Therefore, the dataset could also be used for competitions dedicated to a large set of document analysis problems, including word spotting.