Loading GP-1D model…
Loading Gaussian model…
Loading SIR model…
Loading BO-1D model...
Loading AR-buffer model...
Loading ALINE model...
Interactive amortized probabilistic conditioning in your browser.
ACE, the Amortized Conditioning Engine, is a transformer-based probabilistic model: instead of solving each new inference problem from scratch, it returns the predictions of interest in a single forward pass.
Classically, every new dataset is its own inference problem: you run a computation against that specific data — optimization, sampling, numerical integration — and rerun it when the data changes. Amortized inference moves the cost up front: a neural network is trained on many simulated problem instances; afterwards, solving each new problem is fast. The training effort is amortized over all future uses — hence the name. The trade-off is that answers are approximate, and reliable only within the distribution of problems the network was trained on.
ACE sits in the same broad family as Prior-Fitted Networks (PFNs) and Transformer Neural Processes (TNPs), and is not so different from modern large language models (LLMs): all learn to make predictions by conditioning on a provided set of observations (the context, encoded as tokens). LLMs predict the next word (token) of a text; ACE predicts all sorts of quantities of interest.
ACE extends what a token can be. Observed data, interpretable latent variables, and prior information supplied at runtime all enter the same token set; the model conditions on one set of tokens and predicts distributions over another. Both sets can mix data with task-relevant quantities — model parameters, class labels, the location of an optimum — which lets one model cover a variety of scientific and machine-learning tasks.
Each tab is a small, inspectable task running a compact pretrained network in your browser. Every click, drag, or slider edit poses a fresh conditioning problem — the instant updates are amortization at work. Each tab's "?" button explains the task and the classical computation that the forward pass replaces. The browser models are intentionally small, so expect some imperfections (larger deployed ACE models perform better); warnings appear inside the plots when you leave the training distribution.
This demo is based on the following work from our group and collaborators.
@inproceedings{chang2025amortized,
title={Amortized Probabilistic Conditioning for Optimization, Simulation and Inference},
author={Chang, Paul E and Loka, Nasrulloh and Huang, Daolang and Remes, Ulpu and Kaski, Samuel and Acerbi, Luigi},
booktitle={The Twenty-eighth International Conference on Artificial Intelligence and Statistics (AISTATS 2025)},
year={2025}
}
@inproceedings{hassan2026efficient,
title={Efficient Autoregressive Inference for Transformer Probabilistic Models},
author={Conor Hassan and Nasrulloh Ratu Bagus Satrio Loka and Cen-You Li and Daolang Huang and Paul Edmund Chang and Yang Yang and Francesco Silvestrin and Samuel Kaski and Luigi Acerbi},
year={2026},
booktitle={The Fourteenth International Conference on Learning Representations (ICLR 2026)}
}
@inproceedings{huang2025aline,
title={ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition},
author={Daolang Huang and Xinyi Wen and Ayush Bharti and Samuel Kaski and Luigi Acerbi},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025)},
year={2025}
}
nanoACE is developed by the Machine and Human Intelligence (MHI) group at the University of Helsinki. Work on nanoACE was supported by the Research Council of Finland (Flagship programme: Finnish Center for Artificial Intelligence FCAI; and grants 358980 and 356498) and by the research environment provided by ELLIS Institute Finland.