Searcharxiv⌕ Search

arXiv subjects

Martin Jankowiak

Publications and source records attributed to Martin Jankowiak.

30 records · Page 2Linked to original sources

Pyro: Deep Universal Probabilistic Programming

Pyro is a probabilistic programming language built on Python as a platform for developing advanced probabilistic models in AI research. To scale to large datasets and high-dimensional models, Pyro uses stochastic variational inference algorithms and probability distributions built on top of PyTorch, a modern GPU-accelerated deep learning framework. To accommodate complex or model-specific algorithmic behavior, Pyro leverages Poutine, a library of composable building blocks for modifying the behavior of probabilistic programs.

cs.LG↗

Pathwise Derivatives Beyond the Reparameterization Trick

We observe that gradients computed via the reparameterization trick are in direct correspondence with solutions of the transport equation in the formalism of optimal transport. We use this perspective to compute (approximate) pathwise gradients for probability distributions not directly amenable to the reparameterization trick: Gamma, Beta, and Dirichlet. We further observe that when the reparameterization trick is applied to the Cholesky-factorized multivariate Normal distribution, the resulting gradients are suboptimal in the sense of optimal transport. We derive the optimal gradients and show that they have reduced variance in a Gaussian Process regression task. We demonstrate with a variety of synthetic experiments and stochastic variational inference tasks that our pathwise gradients are competitive with other methods.

stat.ML↗

Uncovering the Spatiotemporal Patterns of Collective Social Activity

Social media users and microbloggers post about a wide variety of (off-line) collective social activities as they participate in them, ranging from concerts and sporting events to political rallies and civil protests. In this context, people who take part in the same collective social activity often post closely related content from nearby locations at similar times, resulting in distinctive spatiotemporal patterns. Can we automatically detect these patterns and thus provide insights into the associated activities? In this paper, we propose a modeling framework for clustering streaming spatiotemporal data, the Spatial Dirichlet Hawkes Process (SDHP), which allows us to automatically uncover a wide variety of spatiotemporal patterns of collective social activity from geolocated online traces. Moreover, we develop an efficient, online inference algorithm based on Sequential Monte Carlo that scales to millions of geolocated posts. Experiments on synthetic data and real data gathered from Twitter show that our framework can recover a wide variety of meaningful social activity patterns in terms of both content and spatiotemporal dynamics, that it yields interesting insights about these patterns, and that it can be used to estimate the location from where a tweet was posted.

cs.SI↗

Jet Substructure Templates: Data-driven QCD Backgrounds for Fat Jet Searches

QCD is often the dominant background to new physics searches for which jet substructure provides a useful handle. Due to the challenges associated with modeling this background, data-driven approaches are necessary. This paper presents a novel method for determining QCD predictions using templates -- probability distribution functions for jet substructure properties as a function of kinematic inputs. Templates can be extracted from a control region and then used to compute background distributions in the signal region. Using Monte Carlo, we illustrate the procedure with two case studies and show that the template approach effectively models the relevant QCD background. This work strongly motivates the application of these techniques to LHC data.

hep-ph↗

Constraining CP-violating Higgs Sectors at the LHC using gluon fusion

We investigate the constraints that the LHC can set on a 126 GeV Higgs boson that is an admixture of CP eigenstates. Traditional analyses rely on Higgs couplings to massive vector bosons, which are suppressed for CP-odd couplings, so that these analyses have limited sensitivity. Instead we focus on Higgs production in gluon fusion, which occurs at the same order in the strong coupling for both CP-even and -odd couplings. We study the Higgs plus two jet final state followed by Higgs decay into a pair of tau leptons. We show that using the 8 TeV dataset it is possible to rule out the pure CP-odd hypothesis in this channel alone at nearly 95\% C.L, assuming that the Higgs is CP-even. We also provide projected limits for the 14 TeV LHC run.

hep-ph↗

Learning How to Count: A High Multiplicity Search for the LHC

We introduce a search technique that is sensitive to a broad class of signals with large final state multiplicities. Events are clustered into large radius jets and jet substructure techniques are used to count the number of subjets within each jet. The search consists of a cut on the total number of subjets in the event as well as the summed jet mass and missing energy. Two different techniques for counting subjets are described and expected sensitivities are presented for eight benchmark signals. These signals exhibit diverse phenomenology, including 2-step cascade decays, direct three body decays, and multi-top final states. We find improved sensitivity to these signals as compared to previous high multiplicity searches as well as a reduced reliance on missing energy requirements. One benefit of this approach is that it allows for natural data driven estimates of the QCD background.

hep-ph↗

LHC probes the hidden sector

In this note we establish LHC limits on a variety of benchmark models for hidden sector physics using 2011 and 2012 data. First, we consider a "hidden" U(1) gauge boson under which all Standard Model particles are uncharged at tree-level and which interacts with the visible sector either via kinetic mixing or higher dimensional operators. Second, we constrain scalar and pseudo-scalar particles interacting with the Standard Model via dimension five operators and Yukawa interactions, in particular including so-called axion-like particles. In both cases we consider several different final states, including photons, electrons, muons and taus, establishing new constraints for a range of GeV to TeV scale masses. Finally, we also comment on particles with electric charges smaller than e that arise from hidden sector matter.

hep-ph↗

Jet Dipolarity: Top Tagging with Color Flow

A new jet observable, dipolarity, is introduced that can distinguish whether a pair of subjets arises from a color singlet source. This observable is incorporated into the HEPTopTagger and is shown to improve discrimination between top jets and QCD jets for moderate to high pT.

hep-ph↗

Angular Scaling in Jets

We introduce a jet shape observable defined for an ensemble of jets in terms of two-particle angular correlations and a resolution parameter R. This quantity is infrared and collinear safe and can be interpreted as a scaling exponent for the angular distribution of mass inside the jet. For small R it is close to the value 2 as a consequence of the approximately scale invariant QCD dynamics. For large R it is sensitive to non-perturbative effects. We describe the use of this correlation function for tests of QCD, for studying underlying event and pile-up effects, and for tuning Monte Carlo event generators.

hep-ph↗

Jet Substructure Without Trees

We present an alternative approach to identifying and characterizing jet substructure. An angular correlation function is introduced that can be used to extract angular and mass scales within a jet without reference to a clustering algorithm. This procedure gives rise to a number of useful jet observables. As an application, we construct a top quark tagging algorithm that is competitive with existing methods.

hep-ph↗

Nearly Supersymmetric Dark Atoms

Theories of dark matter that support bound states are an intriguing possibility for the identity of the missing mass of the Universe. This article proposes a class of models of supersymmetric composite dark matter where the interactions with the Standard Model communicate supersymmetry breaking to the dark sector. In these models supersymmetry breaking can be treated as a perturbation on the spectrum of bound states. Using a general formalism, the spectrum with leading supersymmetry effects is computed without specifying the details of the binding dynamics. The interactions of the composite states with the Standard Model are computed and several benchmark models are described. General features of non-relativistic supersymmetric bound states are emphasized.

hep-ph↗

Experimental constraints on the free fall acceleration of antimatter

In light of recent experimental proposals to measure the free fall acceleration of antihydrogen in the earth's gravitational field, we investigate the bounds that existing experiments place on any asymmetry between the free fall of matter and antimatter. We conclude that existing experiments constrain any such asymmetry to be less than about 10^-7. First we consider contributions to the inertial masses of atoms that encode the presence of antimatter and use precision Eotvos experiments to establish the level at which they satisfy the equivalence principle. In particular we focus on vacuum polarization effects and the antiquark content of nucleons. Second we consider a class of theories that contain long range scalar and vector forces that cancel with one another to some high precision. By construction such theories would be able to evade detection in Eotvos experiments that utilize matter while still allowing for a signal in antimatter experiments. Even taking such cancellation for granted, however, we show that the radiative damping of binary pulsar systems constrains these forces to be significantly weaker than gravity. Furthermore we show that there are limits to the accuracy with which such cancellation can be arranged: first by determining the precision to which scalar charges can track vector charges in the best candidate theories; and, second, by showing that the different velocity dependence of scalar and vector forces necessarily introduces non-cancellation at a quantifiable level.

hep-ph↗