Benchmarking TMVA package against TensorFlow on event-by-event inference performance on multi-layered perceptrons for HEP

HEP has some very specific requirements about the usage of deep learning. Also, HEP is known for outrageous amounts of data produced in a single collision. For example, at LHC we are talking about petabyte per second scales, and with the introduction of HL-LHC, this numbers will grow significantly. Currently, the HEP community wants to find out, is it possible to efficiently run neural networks in low-level triggers, thus reducing the amount of collected data without compromising its quality? This work aims to find answers to this question and the future directions of current deep network tools used in HEP. This work is a technical report on the project I was working between 2nd in July and 24rd of August 2018.