PSILOGO

Laboratory for Particle Physics (LTP)


LTP Colloquium

NeuroBayes: Predictive Analytics for Science and Economy

Thursday, April 10, 2014, 16:00
OSGA/E6

M. Feindt, KIT

Abstract:
The Bayesian Neural Network NeuroBayes was developed 15 years ago for optimizing knowledge discovery from high energy physics experiments at CERN. Over the years there were many successful applications also at Fermilab (USA) and KEK (Japan). The aim is to predict unknown quantities in single events by individualizing statistical statements about it from multivariate measurements obj many events. Recently an artificial intelligence system of 72 NeuroBayes networks performed the equivalent of few hundred PhD theses reconstructing B mesons in more that 1100 decay channels and found more than twice as many events than the whole 400-physicist-collaboration in the ten years of data taking before. NeuroBayes also runs in the trigger of the LHCb-experiment, and a NeuroBayes expert chip is being constructed to perform smart data reduction directly at the sensors in the Belle II experiment where the data rate will be so large as a complete readout to computers is impossible.
The company blue yonder founded by the speaker applies and further develops NeuroBayes for commercial applications in retail, e-commerce, insurance, finance, energy, healthcare and industry. With meanwhile more than 70 PhD physicists in the data science team blue yonder has developed to Europe’s leading predictive analytics company and has been awarded a large number of innovation prizes.