Examining the Effects of Simplified Neural Net Structure on Jet Identification





The summer after my freshman year at Yale, I worked under Tobias Golling in the Yale ATLAS High Energy Research Group. At CERN, particles are accelerated and collided in the Large Hadron Collider (LHC), forming “jets” of subatomic particles. In these collision events, each particle gives off unique signatures, which are measured by calorimeters, spectrometers, and magnets in the detector.

In order to reconstruct the physics of the collision, advanced machine learning techniques such as neural networks are trained to discriminate among subatomic particles, and to “tag” or identify them by their signatures.

While at CERN, I worked to simplify the structure of a neural net used to tag bottom quarks, charm quarks and muons. I examined low level variables, performing principal components analysis to analyze their relative importance, and significantly reduced the number of features used by the network. I removed upstream dependencies and reduced computational intensity while maintaining comparable performance.

Leave a Reply

Your email address will not be published. Required fields are marked *