Bayesian Program Synthesis
In machine learning, Bayesian Program Synthesis (BPS), Bayesian Programs write (synthesize) new Bayesian programs. This is in contrast to the field of probabilistic programs where humans write new probabilistic (Bayesian) programs. Bayesian probabilities is a strategy to learn distributions over Bayesian programs.[1]
Bayesian Program Synthesis can be compared to the work on Bayesian Program Learning by Lake, Salakhutdinov, and Tenenbaum's,[2] where probabilistic program components were hand-written, pre-trained on data, and then hand assembled in order to recognize handwritten characters.[3]
The framework
Bayesian Program Synthesis (BPS) has been described as a framework related to and utilizing probabilistic programming. In BPS, probabilistic programs are generated that are themselves priors over a space of probabilistic programs.[2] This strategy allows more automatic synthesis of new programs via inference and is achieved by the composition of modular component programs.
The modularity in BPS allows inference to work on and test smaller probabilistic programs before being integrated into a larger model.[4]
Bayesian methods and models are frequently used to incorporate prior knowledge. When good prior knowledge can be incorporated into a Bayesian model, effective inference can often be performed with much less data.[5]
This framework can be also be contrasted with the family of automated program synthesis fields, including program synthesis, programming by example, and programming by demonstration. The goal in such fields is to find the best program that satisfies some constraint. In program synthesis, for instance, verification of logical constraints reduce the state space of possible programs, allowing more efficient search to find an optimal program. Bayesian Program Synthesis differs both in that the constraints are probabilistic and the output is itself a distribution over programs that can be further refined.[5]
See also
References
- ↑ Knight, Will. "AI Software Juggles Probabilities to Learn from Less Data" (in en). MIT Technology Review. https://www.technologyreview.com/s/603542/ai-software-juggles-probabilities-to-learn-from-less-data/.
- ↑ 2.0 2.1 Wood, Charlie (2017-02-16). "Startup pairs man with machine to crack the 'black box' of neural networks". Christian Science Monitor. ISSN 0882-7729. http://www.csmonitor.com/Technology/2017/0216/Startup-pairs-man-with-machine-to-crack-the-black-box-of-neural-networks.
- ↑ Lake, Brenden M.; Salakhutdinov, Ruslan; Tenenbaum, Joshua B. (2015-12-11). "Human-level concept learning through probabilistic program induction" (in en). Science 350 (6266): 1332–1338. doi:10.1126/science.aab3050. ISSN 0036-8075. PMID 26659050. http://science.sciencemag.org/content/350/6266/1332.
- ↑ "Talking Machines: Probabilistic programming, with Ben Vigoda | Robohub" (in en-US). http://robohub.org/talking-machines-probabilistic-programming-with-ben-vigoda/.
- ↑ 5.0 5.1 Metz, Cade. "AI’s Factions Get Feisty. But Really, They’re All on the Same Team" (in en-US). WIRED. https://www.wired.com/2017/02/ais-factions-get-feisty-really-theyre-team/.
External links
- Commentary on BPS by David Garrity: Artificial Intelligence to see Significant Progress in 2017
 |
