Gowtham Atluri Mike Sokoloff PV finder LHCb upgrade Much higher pileup Need for faster algorithms Goal Discovering PVs and SVs Use of ML approaches Transform 3D data to 1D 1D Convolutional neural nets ID: 1012631
Download Presentation The PPT/PDF document "Computationally Efficient and accurate M..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. Computationally Efficient and accurate ML algorithms for PV findingGowtham AtluriMike Sokoloff
2. PV finderLHCb upgradeMuch higher pileupNeed for faster algorithmsGoal: Discovering PVs and SVsUse of ML approachesTransform 3D data to 1D1D Convolutional neural netsHighly parallelizable, GPU friendlyRoom for interpretability
3. Hybrid ML approach
4. Generating KDEPartial or full TrackingFill each voxel center with a Gaussian PDFPDF for each track is combinedFill z histogram with maximum KDE in xy.
5. Sample z KDE HistogramChallenges:- True PVs may be offset from peak- Peaks in close proximity interact
6. CNNs developed by Cincinnati LHCb group
7. Opportunities for CS contributionsCharacterize PVs where CNN is accurateIdentify the filters that capture PVsRefine the CNN design for computational efficiencyCharacterize PVs where CNN is not accurateInsufficient training examples?Need for multiple models?Ensemble models?
8. Opportunities for CS contributionsCrosspollination of ideas with wider ML communityHeuristic one-pass algorithmsHidden Markov Models, etc.Mass Spectrometry [Yang et al. 2009]ECG DataAlso in smart watches and mobile devices for step counting.
9. Software deliverablesTools to visualize CNN filtersVisualizing the filters that are learnedVisualizing which filters contribute to correct predictionsVisualizing which filters contribute to incorrect predictionsTools to investigate CNN learningVisualizing filters as a function of epoch#
10. Other OutcomesEvaluation of alternative design choices for CNNsOne full 'z' histogram vs. separate 'chunks'Impact of #layers, size of filters, etc.Use of Recurrent NNs? When PVs are in close proximity.Evaluation of existing ML/heuristic approaches for peak detection Heuristic computational approachesHMMsCNNs
11. Questions?