Inna Kucher

Normalizing flows are generative models that show poor performance on out-of-distribution (OOD) detection tasks with a likelihood-based test. In this study we focus on the "approximate mass" metric. We show that while it improves OOD detection performance, it has limitations under a maximum likelihood training. To solve this limitation we modify the training objective by incorporating the approximate mass. It smooths the learnt distribution in the vicinity of training in-distribution data. We measure an average of 97.6% AUROC in our experiments on different benchmarks, showing an improvement of 16% with respect to the best baseline we tested against.

Modern deep learning architectures, while highly successful, are characterized by substantial computational and memory demands due to their large number of parameters or the storing of activations. That is why it is hard to adapt a neural network to the constraints of hardware, especially at the edge. This paper presents an investigation into a novel approach for activation compression, which we term ’Projection-based compression on channels’ or ’ProChan’. Our method involves interposing projection layers into a pretrained network around the nonlinearity, reducing the channel dimensionality through compression operations and then expanding it back. Our module is made to be then totally fused with the convolutions around it, guaranteeing no overhead, and maximum FLOPs reduction. We studied its absorption of the cost of quantization, to combine the two approaches for footprint reduction. Our findings indicate that the projections likely perform an ’adaptive stretching’ operation on the feature space, enabling the preservation of essential information when constrained by dimensional limitations. We also perform an ablation study on the different possible strategies for a stable and quick training, and analyse the interactions with different quantization paradigms, namely PACT for activations and post-training quantization (PTQ) methods for weights.

Pre-training on an upstream task is widely used in deep learning to boost performance of downstream tasks. Recent studies analyzed pre-training with large datasets and large deep neural network architectures. However, pre-training is very useful in practice when downstream tasks have scarce data and are trained under computational constraints. To assess pre-training performance in this setting, we train different deep architectures with 1M parameters. We create different subsets of ImageNet to study the influence of upstream dataset in detail by varying the total size, but also the ratio between number of classes and samples per class for a constant total size. Then, we use the resulting models in transfer toward six diversified downstream tasks using linear probing and full fine tuning for downstream training. Experimental results confirm previous ones regarding performance saturation in downstream tasks, but we find that saturation occurs faster for compact deep architectures. The use of different ImageNet subsets leads to globally similar performance when enough data is included, regardless of the dataset structure. The comparison of downstream training strategies shows that linear probing can be competitive, particularly for few-shot settings. This is at odds with previous reports, which assert the superiority of full fine tuning. Finally, we observe that the type of deep architecture has a significant effect on results, but that their relative performance varies depending on the downstream training strategy.

The integral observables for model-independent detections of Abelian Z′ gauge boson in e + e - → μ + μ - (τ + τ - ) process with unpolarized beams at the ILC energies are proposed. They are based on the differential cross-section of deviations from the standard model predictions calculated with a low energy effective Lagrangian and taking into consideration the relations between the Z′ couplings to the fermions derived already. Due to these relations, the cross-section exhibits angular distribution giving a possibility for introducing one- or two-parameter observables which effectively fit the mass m Z′ , the axial-vector [Formula: see text] and the product of vector couplings v e v μ (v e v τ ). A discovery reach for the Z′ is estimated for two of introduced observables. Determination of the basic Z′ model is discussed. Comparison with other results and approaches is given.

A search for a Higgs boson decaying to two photons has been performed by the CMS Collaboration at the LHC experiment using pp collisions at a center-of-mass energy of 13TeV with an integrated luminosity of 2.6/fb. In the decay of the Higgs boson into two photons, the unconverted final state photons are not detected in the tracker, so the determination of the associated primary vertex is not trivial. Moreover, the CMS ECAL has no longitudinal segmentation, it is thus not a pointing calorimeter. The information from the recoiling tracks, and, when at least one of the photons is converted in the tracker, from the conversion tracks can be used to determine the primary vertex. The vertex identification algorithm used in the search for the Higgs boson decaying to two photons is described in this presentation, together with its performance. Presented at Moriond/EW2016 51st Rencontres de Moriond on Electroweak Interactions and Unified Theories Vertex identification in the search for the Higgs boson decaying to two photons in CMS Inna Kucher on behalf of the CMS Collaboration CEA Saclay IRFU/SPP (France) A search for the Higgs boson decaying to two photons has been performed by the CMS Collaboration at the LHC experiment using pp collisions at a center-of-mass energy of 13 TeV with an integrated luminosity of 2.7 fb−1. In the decay of the Higgs boson into two photons, the unconverted final state photons are not detected in the tracker. Moreover, the CMS electromagnetic calorimeter (ECAL) has no longitudinal segmentation, it is thus not a pointing calorimeter. As a result, the primary vertex determination of the Higgs boson decaying to two photons is not trivial. The vertex identification algorithm used in the search for the Higgs boson decaying to two photons is described, together with its performance.

A search for massive resonances decaying to a Z boson and a photon is performed in events with a hadronically decaying Z boson candidate, separately in light-quark and b quark decay modes, identified using jet substructure and advanced b tagging techniques. Results are based on samples of proton-proton collisions collected with the CMS detector at the LHC at center-of-mass energies of 8 and 13 TeV, corresponding to integrated luminosities of 19.7 and 2.7 inverse femtobarns, respectively. The results of the search are combined with those of a similar search in the leptonic decay modes of the Z boson, based on the same data sets. Spin-0 resonances with various widths and with masses in a range between 0.2 and 3.0 TeV are considered. No significant excess is observed either in the individual analyses or the combination. The results are presented in terms of upper limits on the production cross section of such resonances and constitute the most stringent limits to date for a wide range of masses.

The integral observables for model-independent detections of Abelian Z' gauge boson in e^+ e^- -> mu^+ mu^- (tau^+ \tau^-) process with unpolarized beams at the ILC energies are proposed. They are based on the differential cross-section of deviations from the standard model predictions calculated with a low energy effective Lagrangian and taking into consideration the relations between the Z' couplings to the fermions derived already. Due to these relations, the cross-section exhibits angular distribution giving a possibility for introducing one- or two parameter observables which effectively fit the mass m_Z', the axial-vector (a_Z)'^2 and the product of vector couplings v_e v_mu (v_e v_tau). Comparison with other results and approaches is given.

In this thesis, the measurement of the Higgs boson properties in the diphoton decay channel with the CMS experiment at the Large Hadron Collider (LHC) is presented. The focus of this work is the tṫH production mode, as it is the only direct access to the top quark Yukawa coupling, a fundamental parameter of the Standard Model. tṫH is a very rare process, two orders of magnitude smaller than the dominant Higgs boson production by gluon fusion. At 13 TeV, ttH production is about 4 times larger than at 8 TeV. This thesis takes over the studies performed at 8 TeV, where the statistics was not enough for an observation of ttH. Despite a very small branching ratio (only about 0.2%), the two photons decay channel of the Higgs boson is very promising, because of its excellent mass resolution (about 1%). Moreover, its signature in the detector is very clear. The diphoton decay channel is also of particular interest as it is the only channel allowing the study of all production modes: gluon fusion, vector boson fusion, associated productions with a W or a Z bosons, or with a top quark pair.The document starts with a theoretical introduction about the Standard Model and Higgs boson physics at LHC, followed by a description of the CMS detector. To achieve an excellent mass resolution in the H → ᵞᵞ channel, the electromagnetic calorimeter has to be calibrated. The laser monitoring system plays an important role in the calibration chain and it is described in details. On the long term, the laser monitoring system will have to be upgraded as level of radiation influences its electronics. I present my work on the possible upgrade of the laser monitoring system, along with the study of its possible precision.H → ᵞᵞ inclusive analysis had several iterations for conferences in 2016 and 2017. The strategy for 2017 is described in this document. An event classification is used to maximize the signal significance and to study specific Higgs boson production modes. My contributions to the H → ᵞᵞ analysis are primary vertex identification, photon identification and the study of the tṫH production mode. Each contribution is described in details in dedicated chapters. The tṫH, H → ᵞᵞ analysis is shown for two iterations in 2016 and 2017, with the emphasis on improvements in 2017 analysis. Finally, the results of the inclusive and tṫH, H → ᵞᵞ analysis, using the full 2016 dataset corresponding to an integrated luminosity of 35.9 fb-1, are shown.

A measurement of the ttbar production cross section at sqrt(s)=13 TeV is presented using proton-proton collisions, corresponding to an integrated luminosity of 2.3 inverse femtobarns, collected with the CMS detector at the LHC. Final states with one isolated charged lepton (electron or muon) and at least one jet are selected and categorized according to the accompanying jet multiplicity. From a likelihood fit to the invariant mass distribution of the isolated lepton and a jet identified as coming from the hadronization of a bottom quark, the cross section is measured to be sigma(ttbar)= 835 +/- 3 (stat) +/- 23 (syst) +/- 23 (lum) pb, in agreement with the standard model prediction. Using the expected dependence of the cross section on the pole mass of the top quark (m[t]), the value of m[t] is found to be 172.7+2.4-2.7 GeV.

Jets produced in hard scattering processes provide an important probe of the thermodynamical and transport properties of the quark gluon plasma (QGP) created in high-energy nuclear (AA) collisions.The products of the hard scattering evolve as parton showers propagating through the medium and experience in-medium energy loss.The information on the initial jet properties can be extracted if an electroweak boson is produced together with parton in the initial hard scattering.Photon and Z boson-tagged measurements of the parton energy loss exploit the fact that the outgoing photon or leptons originating from Z boson are unmodified while traversing the QGP, and thus provide information about the momentum, direction, and flavor of the associated hard-scattered parton before it begins to shower and become quenched.In this proceeding, Z+jet and isolated γ+jet measurements by CMS are reported.

Jets produced in hard scattering processes provide an important probe of the thermodynamical and transport properties of the quark gluon plasma (QGP) created in high-energy nuclear (AA) collisions. The products of the hard scattering evolve as parton showers propagating through the medium and experience in-medium energy loss. The information on the initial jet properties can be extracted if an electroweak boson is produced together with parton in the initial hard scattering. Photon and Z boson-tagged measurements of the parton energy loss exploit the fact that the outgoing photon or leptons originating from Z boson are unmodified while traversing the QGP, and thus provide information about the momentum, direction, and flavor of the associated hard-scattered parton before it begins to shower and become quenched. In this proceeding, Z+jet and isolated $\mathrm{\gamma}$+jet measurements by CMS are reported.