S. Gleyzer - ACAT summary
(incomplete summary here, please also see slides and conference homepage)

• started of as a machine learning conference series in the past and became more general data analysis conference series. Now returned to focus on ML.
• many applications of ML:
  • end-to-end learning (go directly from detector response to event class, w/o reconstruction)
  • GANs for simulation (faster than running full GEANT simulations)
  • track reconstruction (CNN in trigger, RNN for tracking)
  • pile up removal with CNN to correct pile up in jet image
  • flavour tagging
  • trigger applications (lookup table for pT regression BDT, cf. last IML meeting, cat boost)
  • particle identification with XGBoost
• Analysis tools
  • histogrammar by diana: parallelize event loop with spark
  • Root's TDataFrame: use declarative language rather than low level event loop, let framework optimize evaluations, caching, …
  • Root improvements in I/O (compression, parallelization), new TMVA features, web technology for graphics.
• Questions
  • Was there anything on reinforcement learning?
    • A. possibly in one of the posters
• Comment (Sergei): In the summary I left out the GAN for simulation talk by S. Vallecorsa.

T. Golling - Hammers and nails
(incomplete summary here, please also see slides and conference homepage)

• Title "Hammers and nails" refers to Machine Learning as a big hammer technology to tackle various problems, nails being the particle physics problems
• minutes kept on google docs https://docs.google.com/document/d/1y7cE8qVp6xyKdlWzaBqD81_Otvf5YPhHcjlybtURKbU/edit#
• with break out sessions on white/black boards
• ideas for new concepts in HEP:
  • use GAN and VAE to infer w/o likelihood building in between
• Adversarial examples
  • image recognition algorithms can "easily" be tricked into misclassifying an image (without changing it to the human viewer)
  • usable to trick voice recognition to transfer money
  • examples are "crafted", arguable if a detector would spit them out in HEP
• discussion on autonomous driving
  • should AI be taught the rules of the road or learn them just by observation
  • (what about interaction with human drivers and when it's okay to break the rules)
• probabilistic programming
  • goal is to learn a generator which learned its rules.
  • examples of generators which can generate "normal looking trees" or "stable structures"
• Questions:
  • It seems there was very little on classical signal/background classification (Enrico Guiraud)
    • That was a feature of the topic selection for the summary.
  • The classical examples of how to generate adversarial inputs require internal knowledge of the NN that is "attacked" shouldn't that hinder fraud
    • That was discussed, it appears adversarial examples are highly portable in the sense that an adversarial example for one network
    • will also be adversarial for different networks for the same task

G. Kasiecka – Boost summary
(incomplete summary here, please also see slides and conference homepage) 

• Boost is a workshop about advances in jet physics, e.g. heavy object tagging
• See also the machine learning slides in the boost exp. summary talk: https://indico.cern.ch/event/579660/contributions/2496143/attachments/1496921/2330050/Boost17_ExpSum.pdf
• Many talks about classification, but also one each on GANs and pileup
• Classification contributions (see slides for details and references)
  • "color" jet image for quark/gluon jet classifications 
    • "Colors":
      • charged pt
      • neutral pt
      • charged particle multiplicity
    • Deep NN and colors improve significantly at high pt (~ 1 TeV)
    • The method is stable if trained on Herwig or Pythia.
  • q/g tagging in ATLAS
    • Similar technique as in the previous contribution, applied to ATLAS.
    • But some dependence on the MC used for training seen here
  • b-tagging
    • new idea: look at changes in hits multiplicity between subsequent layers in the tracking detector
  • Machine Learning in CMS
    • Comprehensive talk, but one highlight selected for this summary: 
    • the DeepJet architecture
      • multi-class classification using recurrent networks chained to a fully connected one
  • Top-taggers
    • Use constituents 4-vectors, but keeping some of the spirit of the image approach
    • Similar performance to the image approach with calorimeters, but performs better than images with particle flow: no need to pixelate.
  • Boosted top and WW with Atlas
• GAN for jet simulation
  • See also: https://indico.cern.ch/event/595059/contributions/2497383/
  • CaloGAN to simulate calorimeter response
• Pile-up mitigation with Machine learning
  • Color image used to estimate jet variables removing contributed
    • pt of all neutral particles
    • pt of charged from primary vertex
    • pt of charged from secondary vertex
    • Outputs a single channel image: pT of neutral particles from primary vertex
• Questions
  • • Where there news on Boosted Higgs? discussions, but no major update with ML
    • Universal multi-class classification? in progress, mentioned in some contributions
    • What a ML expert with no experience in physics should do? Re-use a publicly available data set is a good way to start
    • Trainable linear combinations in the Deep learning top tagger, why? Different weighted sum can give access to e.g. the mass of the top or reconstruct the W

M. Paganini - CVPR Summary
(incomplete summary here, please also see slides and conference homepage)

• It's mostly a computer vision conference, organized by IEEE, over 4000 participants
• Selected contributions potentially relevant for HEP presented in the summary
• Some interesting trends discussed in slide 5
• DenseNets
  • Resnet: allow to skip any one layer in the network
  • Densenets: 
    • allow to connect any layer with any other layer (not just skipping one)
    • size of the layers grows at each step
    • bottleneck introduced to reduce feature sizes
    • Advantages (See slide 14 for complete list)
      • strong gradients
      • All features learned are used for prediction, simple features in the first layer survive to the last layer.
  • Slide 18: Using Knowledge Graphs for classifications
    • idea is that ML method identifies elements in the images, but ultimate object which we want to identify is not any  one of these particular elements, 
    • if we have a graph which maps the relation between elements and the final class, we may be able to identify the object we want to tag better
    • Example: want to identify some heavy resonance, but our tagger will identify daughters of this object
    • Use knowledge graph to infer class from tagged elements