Probabilistic inference, automated

Automated science needs automated inference.

MetaInference is automating the process of turning experimental data into scientific conclusions.

Experiments can be automated. Getting from the data to a conclusion is still done by hand.

Between a measurement and a scientific claim sits probabilistic inference: fitting models to data, quantifying uncertainty, weighing competing explanations. It is the step that decides what an experiment actually showed, and it is hard to do well.

Today it depends on specialists. Every analysis raises the same questions, and answering them takes expertise that an automated pipeline does not have on hand.

  • aWhich method suits this model and this data?
  • bHow should it be configured and tuned?
  • cHas it converged?
  • dCan the result be trusted?

Four pieces that together make inference a routine, automated part of the scientific pipeline.

  1. i

    Inference inside the scientific pipeline

    The tooling and interfaces that let inference run as a step of an automated pipeline, not a manual stage between steps.

  2. ii

    Agents that make the decisions

    AI agents that make the calls an inference expert makes: which method to run, how to set it up, and how to judge the result.

  3. iii

    Foundation models for amortized inference

    Pretrained models that return a posterior directly, instead of starting from scratch for every new dataset.

  4. iv

    Better inference methods

    More robust algorithms that need less hand-tuning. We also use AI to search the space of inference algorithms itself.

Machine learning researchers from Tübingen.

We have spent years on the problems of Bayesian inference and probabilistic machine learning, including MCMC, sequential Monte Carlo, variational inference and simulation-based inference, and on the open-source software that makes these methods usable.

Working on automated science, or running inference at scale? Write to us.