class: middle, center, title-slide count: false
.large.blue[Matthew Feickert]
.large[(University of Wisconsin-Madison)]
[email protected]
September 12th, 2023
.kol-1-2[ .large[
-
Perspectives from the 8th workshop of the Forum on the interpretation of the LHC results for BSM studies
- First meeting of the LHC EFT Working Group with the Reinterpretation Forum
- Multiple contributions from Belle II, working to follow reinterpretation best practices
-
Much to summarize from an engaging week of discussions with ATLAS, CMS, Belle II, and theory, but focusing on
- Accomplishments in publishing .bold[rich HEPData and full statistical models]
- Work towards publishing and reusing .bold[machine learning workflows] ] ] .kol-1-2[
.kol-1-2[ .huge[
- Rich HEPData archives seen to provide useful information for theory colleagues
- Made very clear that the information that ATLAS is providing is useful and they want to be able to cite it
- Use of ATLAS full statistical models providing enhancement to theory recasting tools by allowing use of control regions during fits and improved combinations ] ] .kol-1-2[
.kol-1-2[ .huge[
- Multiple theory tools are now refactoring to interface with new
speylikelihood building framework - With full statistical models from ATLAS are able to properly handle nuisance parameters for improved combinations ] ] .kol-1-2[
.kol-1-2[ .huge[
- Multiple theory tools are now refactoring to interface with new
speylikelihood building framework - With full statistical models from ATLAS are able to properly handle nuisance parameters for improved combinations ] ] .kol-1-2[
.kol-1-2[ .large[
- The HEP Statistics Serialization Standard (HS3) allows for more publication through HEP universal serialization to JSON
- Goal: Write (model) once, run anywhere
- Current draft version of HS3 available in the
masterbranch of ROOT is being used to serialize a RooFit workspace for ongoing ATLAS analysisBAT.jlimplementation WIPpyhfimplementation planned following future development ] ] .kol-1-2[
.kol-1-2[ .large[
- The HEP Statistics Serialization Standard (HS3) allows for more publication through HEP universal serialization to JSON
- Goal: Write (model) once, run anywhere
- Current draft version of HS3 available in the
masterbranch of ROOT is being used to serialize a RooFit workspace for ongoing ATLAS analysisBAT.jlimplementation WIPpyhfimplementation planned following future development ] ] .kol-1-2[
.kol-1-2[ .huge[
- Growing interest in pursuing full statistical model benefits is evident from CMS studies towards combined ATLAS+CMS top EFT
- Tooling for bidirectional translation from HistFactory to CMS Combine model serialization with goals of simplifying ATLAS+CMS combinations ] ] .kol-1-2[
.caption[Save the EFT: a primer for the ATLAS+CMS combination in the top sector, Kirill Skovpen] ]
.kol-1-3[ .large[
- While full statistical models allow for calculating the profiled likelihood during reinterpretation, the additional computation time might be traded for faster approximations for large surveys
- Learning a DNN representation of the profile likelihood allows for speedup while maintaining sufficient accuracy ] ] .kol-2-3[
.caption[Parametrising profiled likelihoods with neural networks, Humberto Reyes-González]
.large[
- NN model serialzed to ONNX distributed to Zenodo ] ]
.kol-1-2[ .huge[
- Linear grid search in for limit setting in parameter space is computationally intensive
- Instead can learn the exclusion contour with active learning reducing total computing time
- Requires intensive ML workflow though ] ] .kol-1-2[
.kol-1-2[ .large[
- ATL-PHYS-PUB-2023-010 leverages REANA to run the workflows at scale with analysis loop until contour learned ]
.kol-1-2[ .large[
- The "Les Houches guide to reusable ML models in LHC analyses" guidelines (forthcoming) advocates
- ATLAS SUSY has made two ML models public through archival on HEPData via ONNX, though publication of
lwtnnfiles to HEPData has yet to happen
]
.kol-1-2[ .large[
- Supplementary Material
- Improved definitions and documentation of all input and output variables for ML model
- Leverage .bold[validated] analysis code implementation (e.g.
rivetorSimpleAnlaysis) - At minimum, would want a short note that could be uploaded alongside ML model serialization
- ATLAS
SimpleAnalysisand RECAST already puts these within easy reach ] ] .kol-1-2[ .large[ - Validation Material
- Cutflows: When cuts depend on a neural net output
- Along with cutflows need to know exact signal models (provide SHLA files, simulation run cards)
- Input and output plots (especially for important features) ] ]
.huge[
- Clear advantage that ATLAS has over CMS at the moment is publication of full statistical models and rich HEPData
- Broader community loudly wants to use it and wants to cite it when used
- Theory community explicitly requests full statistical models continue to be published
- ML continues to bring new approaches to reinterpretation and community recommendations
- Les Houches guide to reusable ML models in LHC analyses guidelines
- Across theory and experiment advantages being seen in terms of compute saving ]
class: end-slide, center
Backup
- Testing the scalar triplet solution to CDF's heavy $W$ problem at the LHC, Jon Butterworth
- Reduce, Reuse, Reinterpret, Giordon Stark
- Active Learning for analysis reinterpretation and constraining additional physics parameters, Christian Weber
- Using unbinned measurements for new physics, Ben Nachman
- HEP Statistics Serialization Standard, Carsten Burgard
- Reusing Neural Networks: Lessons learned and Suggestions for the future, Tomasz Procter
- Global Effective Field Theory fits from ATLAS, Rahul Balasubramanian
.kol-1-2[ .large[
- Inference-aware binning
- Publish measurement, then for reinterpretation select optimal binning for analysis
- Derivative Measurements
- Allows later for
$f(x)$ when$x$ measured
- Allows later for
- Extension to higher-dimensions
- ML tools extend readily to multiple dimensions ] ] .kol-1-2[
class: end-slide, center count: false
The end.