BNL, Upton, Long Island, New York, USA
The repository of data for the Relativistic Heavy Ion Collider and associated pre-injector accelerators consists of well over half a petabyte of uncompressed data. By todays standard, this is not a large amount of data. However, a large fraction of that data has never been analyzed and likely contains useful information. We will describe in this paper our efforts to use machine learning techniques to pull out new information from existing data. Our focus has been to look at simple problems, such as associating basic statistics on certain data sets and doing predictive analysis on single array data. The tools we have tested include unsupervised learning using Tensorflow, multimode neural networks, hierarchical temporal memory techniques using NuPic, as well as deep learning techniques using Theano and Keras.
