While the most advanced artificial intelligence systems show certain limits, whether they are autonomous vehicles struggling to interpret unusual contexts, or medical tools withdrawn from the due to lack of reliability, another approach, more discreet, arouses a renewed interest: the Probabilistic Programming. This paradigm is not based on massive learning from data, but on the explicit modeling of the world, its uncertainties and its causal relationships.
Unlike deep neural networks, which mainly learn by statistical correlation without formalizing rules, probabilistic models make it possible to directly describe the links between events, causes and effects, according to a logic inspired by Bayesian reasoning. Where the classic approach to Machine Learning seek to reproduce reasons observed, probabilistic programs formulate hypotheses, confront them with observation, and adjust their understanding of the phenomenon.
The differences between these two schools are structural in nature. Deep Learning excels in prediction, but often remains opaque in its reasoning. Conversely, a probabilistic model can produce several explanations compatible with observation, accompanied by a degree of confidence. In the case of seismic surveillance, for example, this type of model can distinguish different possible causes, natural earthquake, industrial explosion, nuclear test, while justifying its hypotheses. This approach was implemented as part of the system of verification of the nuclear test prohibition treaty, where a program developed in a few tens of minutes made it possible to reach a level of analysis which had previously required decades of human expertise.
In the medical field, this ability to reason with little data is also precious. While neural networks require considerable image bases to learn to detect a pathology like melanoma, a probabilistic model can operate from a more limited number of observations, based on pre -existing clinical knowledge. He is not content to classify, he evaluates, explicit, and can recognize the situations where uncertainty remains.
We find similar use cases in logistics. When a hazard affects a supply chain, it is not only a question of reporting a delay, but of understanding the propagation mechanisms, evaluating the scenarios to come and recommend corrective actions. The probabilistic models are here useful for exploring several possible trajectories, taking into account complex and sometimes unpublished dependencies.
These models do not seek to impress. They generate neither images, texts, nor spectacular content. Their strength lies elsewhere: in their ability to model, to explain, to guide the decision. They do not replace the Deep Learning systems, but come to fill a gap, that of structured understanding and explicit uncertainty, often absent from purely statistical approaches.
Today, the probabilistic programming remains a niche field. Less intuitive, more demanding on the technical level, it is still little highlighted in user interfaces and dominant platforms. However, as expectations evolve and the limits of generative AI become more apparent, this approach could gain in relevance, especially in the sectors where rigor and robustness are essential: health, energy, finance, security.
Benchmarks: historical, uses And applications of probabilistic programming
Historical in Short
-
- Years 1990–2000 :: development of the models graphics (networks Bayesians, HMM). The approach stay confined to laboratories Academic.
- 2008 :: launch of language Church by THE Began, first attempt of language probabilistic generalist.
- 2012–2015 :: emergence of languages more accessible as Stan (Columbia University), Pymc3,, Webppl.
- 2017–2020 :: adoption industrial with Pyro (Uber Ai Labs), Edward (Google Brain), Tensorflow Probability.
- From 2021 :: climb in power of Numpyro,, base on Jax, For of the models Bayesian rapid And differentiables.
Actors And companies pioneer
-
- Uber used Pyro In his models of forecast of request, of pricing dynamic And of planning logistics.
- Google has developed Edward Then there integrated In Tensorflow Probability For there modeling uncertain.
- Microsoft has designed Infer.NET,, used notably In his products Office (as Outlook) For infer automatically of the priorities.
- NASA is based on of the models probabilists For anticipate THE failures In THE systems embedded.
- Governments And institutions :: THE Ctbto (Understanding Nuclear-Test-Ban Treaty Organization) uses of the models probabilists For analysis of the signals seismic.
Examples concrete application
-
- Detection testing nuclear :: A model probabilistic has summer used For distinguish in a few minutes A earthquake of a essay underground, surpassing of the methods classic used during of the decades.
- Medicine personalized :: of the models as Bayesrx are tested For estimate THE reactions individual has some treatments in oncology.
- Diagnosis of breakdown In industry :: some platforms of maintenance predictive model THE relations causal between signals sensors And incidents technical.
- Forecast of the disruption logistics :: of the systems probabilists are used For simulate of the chains supply below uncertainty And to propose of the plans of resilience.
- Security computer science :: For detect of the behavior unusual on of the networks, some Siem (Security Information and Event Management) integrate of the models Bayesian hierarchical.
