The document discusses using a convolutional neural network (CNN) to quickly find primary vertices (PVs) in high-energy physics events recorded by the LHCb experiment. A prototype tracking algorithm is used to generate a 1D kernel density estimate (KDE) histogram from hit triplets. This histogram is then used to train a CNN to predict the locations of PVs. Initial results show the CNN approach can find PVs with 70-75% efficiency and a false positive rate of 0.08-0.13, outperforming current algorithms. Further work aims to improve resolution, find secondary vertices, and integrate the approach into iterative tracking.
HOW 2019: Machine Learning for the Primary Vertex ReconstructionHenry Schreiner
The document describes a machine learning approach for primary vertex reconstruction in high-energy physics experiments. A hybrid method is proposed that uses a 1D convolutional neural network to analyze histograms produced from tracking data. The network is able to find primary vertices with high efficiency and tunable false positive rates, demonstrating the potential of machine learning for this task. Future work involves adding more tracking information and iterating between track association and vertex finding to improve performance.
ACAT 2019: A hybrid deep learning approach to vertexingHenry Schreiner
This document presents a hybrid deep learning approach for vertex finding in high-energy physics experiments. It uses a 1D convolutional neural network to analyze kernel density estimates of track information in order to identify primary vertex positions. The approach achieves primary vertex finding efficiencies of 88-94% with low false positive rates comparable to traditional algorithms. The authors demonstrate tuning of the efficiency-false positive rate tradeoff and discuss plans to improve performance by incorporating additional track information and iterative refinement.
2019 IML workshop: A hybrid deep learning approach to vertexingHenry Schreiner
A hybrid deep learning approach is proposed for vertex finding using 1D convolutional neural networks on kernel density estimates from tracking data. The approach generates 1D histograms from 3D tracking data and uses a CNN to classify primary vertex positions. In a proof-of-concept on simulated data, it achieves primary vertex finding efficiencies and false positive rates comparable to traditional algorithms, with tunable efficiency-false positive tradeoffs. Future work includes incorporating additional tracking features, associating tracks to vertices, and deploying the inference engine for the LHCb trigger.
2019 CtD: A hybrid deep learning approach to vertexingHenry Schreiner
This document presents a hybrid deep learning approach for vertex finding using 1D convolutional neural networks. It describes generating 1D kernel densities from tracking information, building target distributions, and using a CNN architecture with an adjustable cost function to optimize the false positive rate versus efficiency. The approach achieves 93.87% efficiency with a 0.251 false positive rate on test data. Future work includes incorporating additional xy information and exploring full 2D kernel densities.
Jonathan Lefman presents his work on Superresolution chemical microscopyJonathan Lefman
This document discusses several microscopy techniques including structured illumination fluorescence microscopy, time-of-flight secondary ion mass spectrometry, coherent anti-Stokes Raman scattering microscopy, photoactivated localization microscopy, stimulated emission depletion microscopy, and 4Pi microscopy. It focuses on describing improvements made to structured illumination fluorescence microscopy including parallel GPU processing to accelerate image analysis and a new automated imaging framework. Time-of-flight secondary ion mass spectrometry imaging is discussed with applications to iterative clustering and classification analysis.
Adaptive Channel Prediction, Beamforming and Scheduling Design for 5G V2I Net...T. E. BOGALE
The document proposes and evaluates an adaptive channel prediction, beamforming, and scheduling design for 5G vehicle-to-infrastructure networks. It presents an RLS-based algorithm to predict time-varying channel impulse responses and jointly optimizes beamforming vectors and vehicle scheduling to maximize throughput. Simulation results show the proposed design outperforms alternatives when scheduling a single vehicle, but performance degrades with increasing numbers of scheduled vehicles due to accumulated prediction errors.
Artificial Neural Networks for Storm Surge Prediction in North CarolinaAnton Bezuglov
Feedforward Artificial Neural network (FF ANN) for storm surge prediction in North Carolina. Presentation at Coastal Resilience Center by Anton Bezuglov, Ph.D. Usage of TensorFlow and Python with links to the code on GitHub.
Hashing has witnessed an increase in popularity over the
past few years due to the promise of compact encoding and fast query
time. In order to be effective hashing methods must maximally preserve
the similarity between the data points in the underlying binary representation.
The current best performing hashing techniques have utilised
supervision. In this paper we propose a two-step iterative scheme, Graph
Regularised Hashing (GRH), for incrementally adjusting the positioning
of the hashing hypersurfaces to better conform to the supervisory signal:
in the first step the binary bits are regularised using a data similarity
graph so that similar data points receive similar bits. In the second
step the regularised hashcodes form targets for a set of binary classifiers
which shift the position of each hypersurface so as to separate opposite
bits with maximum margin. GRH exhibits superior retrieval accuracy to
competing hashing methods.
HOW 2019: Machine Learning for the Primary Vertex ReconstructionHenry Schreiner
The document describes a machine learning approach for primary vertex reconstruction in high-energy physics experiments. A hybrid method is proposed that uses a 1D convolutional neural network to analyze histograms produced from tracking data. The network is able to find primary vertices with high efficiency and tunable false positive rates, demonstrating the potential of machine learning for this task. Future work involves adding more tracking information and iterating between track association and vertex finding to improve performance.
ACAT 2019: A hybrid deep learning approach to vertexingHenry Schreiner
This document presents a hybrid deep learning approach for vertex finding in high-energy physics experiments. It uses a 1D convolutional neural network to analyze kernel density estimates of track information in order to identify primary vertex positions. The approach achieves primary vertex finding efficiencies of 88-94% with low false positive rates comparable to traditional algorithms. The authors demonstrate tuning of the efficiency-false positive rate tradeoff and discuss plans to improve performance by incorporating additional track information and iterative refinement.
2019 IML workshop: A hybrid deep learning approach to vertexingHenry Schreiner
A hybrid deep learning approach is proposed for vertex finding using 1D convolutional neural networks on kernel density estimates from tracking data. The approach generates 1D histograms from 3D tracking data and uses a CNN to classify primary vertex positions. In a proof-of-concept on simulated data, it achieves primary vertex finding efficiencies and false positive rates comparable to traditional algorithms, with tunable efficiency-false positive tradeoffs. Future work includes incorporating additional tracking features, associating tracks to vertices, and deploying the inference engine for the LHCb trigger.
2019 CtD: A hybrid deep learning approach to vertexingHenry Schreiner
This document presents a hybrid deep learning approach for vertex finding using 1D convolutional neural networks. It describes generating 1D kernel densities from tracking information, building target distributions, and using a CNN architecture with an adjustable cost function to optimize the false positive rate versus efficiency. The approach achieves 93.87% efficiency with a 0.251 false positive rate on test data. Future work includes incorporating additional xy information and exploring full 2D kernel densities.
Jonathan Lefman presents his work on Superresolution chemical microscopyJonathan Lefman
This document discusses several microscopy techniques including structured illumination fluorescence microscopy, time-of-flight secondary ion mass spectrometry, coherent anti-Stokes Raman scattering microscopy, photoactivated localization microscopy, stimulated emission depletion microscopy, and 4Pi microscopy. It focuses on describing improvements made to structured illumination fluorescence microscopy including parallel GPU processing to accelerate image analysis and a new automated imaging framework. Time-of-flight secondary ion mass spectrometry imaging is discussed with applications to iterative clustering and classification analysis.
Adaptive Channel Prediction, Beamforming and Scheduling Design for 5G V2I Net...T. E. BOGALE
The document proposes and evaluates an adaptive channel prediction, beamforming, and scheduling design for 5G vehicle-to-infrastructure networks. It presents an RLS-based algorithm to predict time-varying channel impulse responses and jointly optimizes beamforming vectors and vehicle scheduling to maximize throughput. Simulation results show the proposed design outperforms alternatives when scheduling a single vehicle, but performance degrades with increasing numbers of scheduled vehicles due to accumulated prediction errors.
Artificial Neural Networks for Storm Surge Prediction in North CarolinaAnton Bezuglov
Feedforward Artificial Neural network (FF ANN) for storm surge prediction in North Carolina. Presentation at Coastal Resilience Center by Anton Bezuglov, Ph.D. Usage of TensorFlow and Python with links to the code on GitHub.
Hashing has witnessed an increase in popularity over the
past few years due to the promise of compact encoding and fast query
time. In order to be effective hashing methods must maximally preserve
the similarity between the data points in the underlying binary representation.
The current best performing hashing techniques have utilised
supervision. In this paper we propose a two-step iterative scheme, Graph
Regularised Hashing (GRH), for incrementally adjusting the positioning
of the hashing hypersurfaces to better conform to the supervisory signal:
in the first step the binary bits are regularised using a data similarity
graph so that similar data points receive similar bits. In the second
step the regularised hashcodes form targets for a set of binary classifiers
which shift the position of each hypersurface so as to separate opposite
bits with maximum margin. GRH exhibits superior retrieval accuracy to
competing hashing methods.
Efficient Data Stream Classification via Probabilistic Adaptive WindowsAlbert Bifet
This document discusses efficient data stream classification using probabilistic adaptive windows. It introduces the concept of data streams which have potentially infinite sequences of high-speed data that must be processed in real-time with limited memory. It then describes the probabilistic approximate window (PAW) algorithm, which maintains a sample of data instances in logarithmic memory by giving greater weight to newer instances. The document evaluates several data stream classification methods on real and synthetic data streams and finds that k-nearest neighbors with PAW has higher accuracy and lower memory usage than other methods.
1) The document describes a real-time GPU implementation of visual smoke simulation using the incompressible Navier-Stokes equations.
2) Key steps in the simulation algorithm include adding forces, advecting velocity and scalar fields, solving for pressure, projecting the velocity field, and applying boundary conditions.
3) Volume rendering is achieved by slicing the 3D grid from the viewer's perspective and compositing the slices using the "under" operator, implementing shadows using half-angle slicing.
Neighbourhood Preserving Quantisation for LSH SIGIR PosterSean Moran
This document proposes a neighbourhood preserving quantisation (NPQ) method for locality sensitive hashing (LSH) that assigns multiple bits per hyperplane using multiple thresholds, rather than the standard single bit. The NPQ method optimizes an F1 score using pairwise constraints from training data to determine threshold values. Evaluation on image retrieval tasks shows NPQ consistently outperforms single and double bit baselines across different projection methods, achieving higher precision-recall curves, especially at higher bit rates. Future work includes exploring variable bits per hyperplane and full retrieval evaluations.
This document discusses GPU-based raycasting of volumetric data. It presents an adaptive sampling raycasting algorithm for layered grid data and compares its results to a traditional algorithm. The adaptive algorithm samples rays non-uniformly based on intersecting cell boundaries to more efficiently render layered ocean and atmosphere data on the GPU. Results show the adaptive method produces similar images to the original with fewer samples. Future work involves applying this approach to curvilinear grids and direct volume rendering.
This document discusses a supercomputer called HYPE-2 built by Santosh Pandey, Ram Sharan Chaulagain, and Prakash Gyawali under the supervision of Prof. Dr. Subarna Shakya. It provides an overview of multiprocessor and multicore systems and discusses how HYPE-2 uses a distributed memory architecture with dynamic scaling to achieve high performance computing capabilities for research applications like cryptography, data mining, and weather forecasting. Performance tests showed near-linear speedup as nodes were added, with the system able to handle complex computations through inter-process communication, though it is not as powerful as larger supercomputers.
Fast Perceptron Decision Tree Learning from Evolving Data StreamsAlbert Bifet
The document proposes using perceptron learners at the leaves of Hoeffding decision trees to improve performance on data streams. It introduces a new evaluation metric called RAM-Hours that considers both time and memory usage. The authors empirically evaluate different classifier models, including Hoeffding trees with perceptron and naive Bayes learners at leaves, on several datasets. Results show that hybrid models like Hoeffding naive Bayes perceptron trees often provide the best balance of accuracy, time and memory usage.
Graph500 and Green Graph500 benchmarks on SGI UV2000 @ SGI UG SC14Yuichiro Yasui
The document discusses Graph500 and Green Graph500 benchmarks for evaluating graph processing performance on the SGI UV2000 system. It provides an overview of the benchmarks and describes testing various graph workloads, including social networks and road networks, on different hardware from smartphones to supercomputers. The authors aim to optimize breadth-first search (BFS) graph algorithms on the NUMA-based SGI UV2000 without using MPI through NUMA-aware techniques.
NUMA-aware Scalable Graph Traversal on SGI UV SystemsYuichiro Yasui
The document discusses NUMA-aware scalable graph traversal on SGI UV systems. It proposes an efficient NUMA-aware breadth-first search (BFS) algorithm for large-scale graph processing by pruning remote edge traversals. Numerical results on SGI UV 300 systems with 32 sockets show the algorithm achieves 219 billion traversed edges per second (GTEPS), setting a new single-node performance record on the Graph500 benchmark.
This document discusses processing large time-of-flight secondary ion mass spectrometry (ToF-SIMS) datasets. ToF-SIMS spectrometers can generate hyperspectral image datasets containing millions of voxels and spectral channels. Non-negative matrix factorization (NMF) was performed on an unbinned ToF-SIMS dataset containing over 8 trillion data points from a fingerprint contamination sample. Subsampling was used to address memory limitations. NMF separated fingerprint components from silicon peaks and identified systematic peak misalignments. MapReduce processing is proposed for even larger datasets to distribute computations across nodes.
GoogLeNet introduced several key insights for designing efficient deep learning networks:
1. Exploit local correlations in images by concatenating 1x1, 3x3, and 5x5 convolutions along with pooling.
2. Decrease dimensions before expensive convolutions using 1x1 convolutions for dimension reduction.
3. Stack inception modules upon each other, occasionally inserting max pooling layers, to allow tweaking each module.
4. Counter vanishing gradients with intermediate losses added to the total loss for training deep networks.
5. End with a global average pooling layer instead of fully connected layers to avoid overfitting.
1) The document discusses using an autocorrelation function (ACF) filter on burst image sequences to reduce noise in CMOS image sensors and achieve high quality imaging.
2) It explains that the ACF calculates correlation values based on pixel values sampled over time to distinguish random noise from true signals. Noise pixels will have lower ACF values while true signals have higher values near 1.
3) The algorithm judges each pixel, applying a leveling filter only to pixels below thresholds for value and ACF. This reduces random noise without impacting bright pixels and resolution. Results show noise reduction while maintaining detail.
FDTD Analysis of the Complex Current Distribution on a circular disk exposed ...kagikenco
This document summarizes research on using FDTD analysis to simulate the complex current distribution on a circular disk exposed to an obliquely incident plane electromagnetic wave. It finds that FDTD simulation can accurately reproduce the rigorous solution for current distribution if the incident wave is properly modeled. Specifically, expanding the plane wave source to prevent it from changing to a spherical wave at oblique incidence improves the accuracy of the simulation results. This validated FDTD method can then be effectively used to design parabolic antennas with similar geometries to the circular disk.
Using Very High Resolution Satellite Images for Planning Activities in MiningArgongra Gis
Pleiades satellite imagery can be used to generate high-resolution digital elevation models, contour lines, and 3D models for mining sector planning activities. A case study acquired 240 sq km of Pleiades stereo imagery with less than 5% cloud cover to generate a 5m DEM and detailed contour lines and 3D visualizations. The Pleiades data provided more accurate topographic information than existing SRTM data for environmental studies, monitoring mining volume changes, and other planning purposes in the mining sector.
HACKSing heterogeneity in cell motilityHee June Choi
This document summarizes a study that used live cell imaging and time series clustering to analyze heterogeneity in cell motility at the subcellular level. The study developed a method called HACKS to extract local velocity and fluorescence intensity time series from imaging data. Time series clustering identified distinct protrusion phenotypes ("fluctuating", "periodic", "accelerating"). Molecular dynamics analysis associated the "accelerating" phenotype with temporally ordered recruitment of the actin regulator VASP. Drug inhibition experiments confirmed VASP promotes the "accelerating" protrusion phenotype.
This document proposes a generalized division-free architecture and compact memory structure for resampling in particle filters. It aims to avoid the high hardware cost of traditional multinomial resampling by using accumulators and comparators instead of division and normalization. The architecture is independent of the number of particles and can be used for different resampling methods. Memory usage is optimized by accumulating weights and random numbers on-the-fly instead of storing cumulative sums, reducing area by up to 45% and memory usage by up to 50%. The architecture achieves resampling without ordering, normalization or generating ordered random numbers.
Сегментация объектов на спутниковых снимках (Kaggle DSTL) / Артур Кузин (Avito)Ontico
РИТ++ 2017, секция ML + IoT + ИБ
Зал Белу-Оризонти, 6 июня, 16:00
Тезисы:
http://ritfest.ru/2017/abstracts/2802.html
В докладе я расскажу про решение задачи сегментации объектов на спутниковых снимках, которая была поставлена в рамках Kaggle-соревнования Dstl Satellite Imagery Feature Detection. В этом соревновании я в команде с Романом Соловьёвым занял 2 место.
В докладе я кратко опишу особенности работы нейросети для сегментации объектов. Затем будут показаны примеры модификаций нейросети с учетом особенности задачи. Также будут рассказаны приемы обучения нейросети, значимо повышающие финальную точность. Будут рассказаны все топ-5 решения.
В качестве бонуса - история, как можно сломать лидерборд за пару дней до конца соревнования.
STRIP: stream learning of influence probabilities.Albert Bifet
This document presents a method called STRIP (Streaming Learning of Influence Probabilities) for learning influence probabilities between users in a social network from a streaming log of propagations. It describes three solutions: (1) storing the whole social graph in memory, (2) using min-wise independent hashing to estimate probabilities while using sublinear space, and (3) estimating probabilities only for the most active users to be more space efficient. Experimental results on a Twitter dataset showed these solutions provided good approximations while using reasonable memory and processing time.
Fast R-CNN is a method that improves object detection speed and accuracy over previous methods like R-CNN and SPPnet. It uses a region of interest pooling layer and multi-task loss to jointly train a convolutional neural network for classification and bounding box regression in a single stage of training. This allows the entire network to be fine-tuned end-to-end for object detection, resulting in faster training and testing compared to previous methods while achieving state-of-the-art accuracy on standard datasets. Specifically, Fast R-CNN trains 9x faster than R-CNN and runs 200x faster at test time.
1) The document discusses combining remote sensing data and in situ monitoring networks to detect extreme events and determine optimal network size and design.
2) It defines ecosystem extreme events using remote sensing proxies and assesses how well existing networks detect extremes using a theoretical model.
3) The analysis finds that detection probabilities of extreme events scale with network size, and that systematically designed networks like ICOS and NEON outperform more randomly distributed networks like Fluxnet for detecting extremes.
Efficient Data Stream Classification via Probabilistic Adaptive WindowsAlbert Bifet
This document discusses efficient data stream classification using probabilistic adaptive windows. It introduces the concept of data streams which have potentially infinite sequences of high-speed data that must be processed in real-time with limited memory. It then describes the probabilistic approximate window (PAW) algorithm, which maintains a sample of data instances in logarithmic memory by giving greater weight to newer instances. The document evaluates several data stream classification methods on real and synthetic data streams and finds that k-nearest neighbors with PAW has higher accuracy and lower memory usage than other methods.
1) The document describes a real-time GPU implementation of visual smoke simulation using the incompressible Navier-Stokes equations.
2) Key steps in the simulation algorithm include adding forces, advecting velocity and scalar fields, solving for pressure, projecting the velocity field, and applying boundary conditions.
3) Volume rendering is achieved by slicing the 3D grid from the viewer's perspective and compositing the slices using the "under" operator, implementing shadows using half-angle slicing.
Neighbourhood Preserving Quantisation for LSH SIGIR PosterSean Moran
This document proposes a neighbourhood preserving quantisation (NPQ) method for locality sensitive hashing (LSH) that assigns multiple bits per hyperplane using multiple thresholds, rather than the standard single bit. The NPQ method optimizes an F1 score using pairwise constraints from training data to determine threshold values. Evaluation on image retrieval tasks shows NPQ consistently outperforms single and double bit baselines across different projection methods, achieving higher precision-recall curves, especially at higher bit rates. Future work includes exploring variable bits per hyperplane and full retrieval evaluations.
This document discusses GPU-based raycasting of volumetric data. It presents an adaptive sampling raycasting algorithm for layered grid data and compares its results to a traditional algorithm. The adaptive algorithm samples rays non-uniformly based on intersecting cell boundaries to more efficiently render layered ocean and atmosphere data on the GPU. Results show the adaptive method produces similar images to the original with fewer samples. Future work involves applying this approach to curvilinear grids and direct volume rendering.
This document discusses a supercomputer called HYPE-2 built by Santosh Pandey, Ram Sharan Chaulagain, and Prakash Gyawali under the supervision of Prof. Dr. Subarna Shakya. It provides an overview of multiprocessor and multicore systems and discusses how HYPE-2 uses a distributed memory architecture with dynamic scaling to achieve high performance computing capabilities for research applications like cryptography, data mining, and weather forecasting. Performance tests showed near-linear speedup as nodes were added, with the system able to handle complex computations through inter-process communication, though it is not as powerful as larger supercomputers.
Fast Perceptron Decision Tree Learning from Evolving Data StreamsAlbert Bifet
The document proposes using perceptron learners at the leaves of Hoeffding decision trees to improve performance on data streams. It introduces a new evaluation metric called RAM-Hours that considers both time and memory usage. The authors empirically evaluate different classifier models, including Hoeffding trees with perceptron and naive Bayes learners at leaves, on several datasets. Results show that hybrid models like Hoeffding naive Bayes perceptron trees often provide the best balance of accuracy, time and memory usage.
Graph500 and Green Graph500 benchmarks on SGI UV2000 @ SGI UG SC14Yuichiro Yasui
The document discusses Graph500 and Green Graph500 benchmarks for evaluating graph processing performance on the SGI UV2000 system. It provides an overview of the benchmarks and describes testing various graph workloads, including social networks and road networks, on different hardware from smartphones to supercomputers. The authors aim to optimize breadth-first search (BFS) graph algorithms on the NUMA-based SGI UV2000 without using MPI through NUMA-aware techniques.
NUMA-aware Scalable Graph Traversal on SGI UV SystemsYuichiro Yasui
The document discusses NUMA-aware scalable graph traversal on SGI UV systems. It proposes an efficient NUMA-aware breadth-first search (BFS) algorithm for large-scale graph processing by pruning remote edge traversals. Numerical results on SGI UV 300 systems with 32 sockets show the algorithm achieves 219 billion traversed edges per second (GTEPS), setting a new single-node performance record on the Graph500 benchmark.
This document discusses processing large time-of-flight secondary ion mass spectrometry (ToF-SIMS) datasets. ToF-SIMS spectrometers can generate hyperspectral image datasets containing millions of voxels and spectral channels. Non-negative matrix factorization (NMF) was performed on an unbinned ToF-SIMS dataset containing over 8 trillion data points from a fingerprint contamination sample. Subsampling was used to address memory limitations. NMF separated fingerprint components from silicon peaks and identified systematic peak misalignments. MapReduce processing is proposed for even larger datasets to distribute computations across nodes.
GoogLeNet introduced several key insights for designing efficient deep learning networks:
1. Exploit local correlations in images by concatenating 1x1, 3x3, and 5x5 convolutions along with pooling.
2. Decrease dimensions before expensive convolutions using 1x1 convolutions for dimension reduction.
3. Stack inception modules upon each other, occasionally inserting max pooling layers, to allow tweaking each module.
4. Counter vanishing gradients with intermediate losses added to the total loss for training deep networks.
5. End with a global average pooling layer instead of fully connected layers to avoid overfitting.
1) The document discusses using an autocorrelation function (ACF) filter on burst image sequences to reduce noise in CMOS image sensors and achieve high quality imaging.
2) It explains that the ACF calculates correlation values based on pixel values sampled over time to distinguish random noise from true signals. Noise pixels will have lower ACF values while true signals have higher values near 1.
3) The algorithm judges each pixel, applying a leveling filter only to pixels below thresholds for value and ACF. This reduces random noise without impacting bright pixels and resolution. Results show noise reduction while maintaining detail.
FDTD Analysis of the Complex Current Distribution on a circular disk exposed ...kagikenco
This document summarizes research on using FDTD analysis to simulate the complex current distribution on a circular disk exposed to an obliquely incident plane electromagnetic wave. It finds that FDTD simulation can accurately reproduce the rigorous solution for current distribution if the incident wave is properly modeled. Specifically, expanding the plane wave source to prevent it from changing to a spherical wave at oblique incidence improves the accuracy of the simulation results. This validated FDTD method can then be effectively used to design parabolic antennas with similar geometries to the circular disk.
Using Very High Resolution Satellite Images for Planning Activities in MiningArgongra Gis
Pleiades satellite imagery can be used to generate high-resolution digital elevation models, contour lines, and 3D models for mining sector planning activities. A case study acquired 240 sq km of Pleiades stereo imagery with less than 5% cloud cover to generate a 5m DEM and detailed contour lines and 3D visualizations. The Pleiades data provided more accurate topographic information than existing SRTM data for environmental studies, monitoring mining volume changes, and other planning purposes in the mining sector.
HACKSing heterogeneity in cell motilityHee June Choi
This document summarizes a study that used live cell imaging and time series clustering to analyze heterogeneity in cell motility at the subcellular level. The study developed a method called HACKS to extract local velocity and fluorescence intensity time series from imaging data. Time series clustering identified distinct protrusion phenotypes ("fluctuating", "periodic", "accelerating"). Molecular dynamics analysis associated the "accelerating" phenotype with temporally ordered recruitment of the actin regulator VASP. Drug inhibition experiments confirmed VASP promotes the "accelerating" protrusion phenotype.
This document proposes a generalized division-free architecture and compact memory structure for resampling in particle filters. It aims to avoid the high hardware cost of traditional multinomial resampling by using accumulators and comparators instead of division and normalization. The architecture is independent of the number of particles and can be used for different resampling methods. Memory usage is optimized by accumulating weights and random numbers on-the-fly instead of storing cumulative sums, reducing area by up to 45% and memory usage by up to 50%. The architecture achieves resampling without ordering, normalization or generating ordered random numbers.
Сегментация объектов на спутниковых снимках (Kaggle DSTL) / Артур Кузин (Avito)Ontico
РИТ++ 2017, секция ML + IoT + ИБ
Зал Белу-Оризонти, 6 июня, 16:00
Тезисы:
http://ritfest.ru/2017/abstracts/2802.html
В докладе я расскажу про решение задачи сегментации объектов на спутниковых снимках, которая была поставлена в рамках Kaggle-соревнования Dstl Satellite Imagery Feature Detection. В этом соревновании я в команде с Романом Соловьёвым занял 2 место.
В докладе я кратко опишу особенности работы нейросети для сегментации объектов. Затем будут показаны примеры модификаций нейросети с учетом особенности задачи. Также будут рассказаны приемы обучения нейросети, значимо повышающие финальную точность. Будут рассказаны все топ-5 решения.
В качестве бонуса - история, как можно сломать лидерборд за пару дней до конца соревнования.
STRIP: stream learning of influence probabilities.Albert Bifet
This document presents a method called STRIP (Streaming Learning of Influence Probabilities) for learning influence probabilities between users in a social network from a streaming log of propagations. It describes three solutions: (1) storing the whole social graph in memory, (2) using min-wise independent hashing to estimate probabilities while using sublinear space, and (3) estimating probabilities only for the most active users to be more space efficient. Experimental results on a Twitter dataset showed these solutions provided good approximations while using reasonable memory and processing time.
Fast R-CNN is a method that improves object detection speed and accuracy over previous methods like R-CNN and SPPnet. It uses a region of interest pooling layer and multi-task loss to jointly train a convolutional neural network for classification and bounding box regression in a single stage of training. This allows the entire network to be fine-tuned end-to-end for object detection, resulting in faster training and testing compared to previous methods while achieving state-of-the-art accuracy on standard datasets. Specifically, Fast R-CNN trains 9x faster than R-CNN and runs 200x faster at test time.
1) The document discusses combining remote sensing data and in situ monitoring networks to detect extreme events and determine optimal network size and design.
2) It defines ecosystem extreme events using remote sensing proxies and assesses how well existing networks detect extremes using a theoretical model.
3) The analysis finds that detection probabilities of extreme events scale with network size, and that systematically designed networks like ICOS and NEON outperform more randomly distributed networks like Fluxnet for detecting extremes.
Implementation of the fully adaptive radar framework: Practical limitationsLuis Úbeda Medina
The document discusses the practical limitations of implementing a fully adaptive radar framework (FAR). It begins by outlining the key components of the FAR, including how the sensor parameters can be adaptively changed by a controller to better fit the system's needs based on information about the environment. It then presents the notation used and provides an example use case of tracking a target moving in a 2D environment using a sensor network with limited resources. Finally, it states that the last section will discuss the practical limitations of the FAR framework.
The document discusses various clustering techniques. It begins by introducing clustering and some common techniques, including partitioning algorithms like k-means, hierarchical algorithms, and density-based algorithms. It then focuses on explaining the k-means algorithm in detail, providing pseudocode and an example application with 16 data points. Key aspects of k-means discussed include initializing cluster centroids, calculating distances between data points and centroids, assigning points to clusters, updating centroids, and determining convergence. The document concludes by analyzing pros and cons of k-means, such as the need to specify the number of clusters k and properly select initial centroids to avoid local optima.
A deep learning model using convolutional neural networks is proposed for lithography hotspot detection. The model takes layout clip images as input and outputs a prediction of hotspot or non-hotspot. It uses several convolutional and pooling layers to automatically learn features from the images without manual feature engineering. Evaluation shows the deep learning model achieves higher accuracy than previous shallow learning methods that rely on manually designed features.
Scratch to Supercomputers: Bottoms-up Build of Large-scale Computational Lens...inside-BigData.com
In this deck from the 2018 Swiss HPC Conference, Gilles Fourestey from EPFL presents: Scratch to Supercomputers: Bottoms-up Build of Large-scale Computational Lensing Software.
"LENSTOOL is a gravitational lensing software that models mass distribution of galaxies and clusters. It was developed by Prof. Kneib, head of the LASTRO lab at EPFL, et al., starting from 1996. It is used to obtain sub-percent precision measurements of the total mass in galaxy clusters and constrain the dark matter self-interaction cross-section, a crucial ingredient to understanding its nature.
However, LENSTOOL lacks efficient vectorization and only uses OpenMP, which limits its execution to one node and can lead to execution times that exceed several months. Therefore, the LASTRO and the EPFL HPC group decided to rewrite the code from scratch and in order to minimize risk and maximize performance, a bottom-up approach that focuses on exposing parallelism at hardware and instruction levels was used. The result is a high performance code, fully vectorized on Xeon, Xeon Phis and GPUs that currently scales up to hundreds of nodes on CSCS’ Piz Daint, one of the fastest supercomputers in the world."
Watch the video: https://wp.me/p3RLHQ-ili
Learn more: https://infoscience.epfl.ch/record/234382/files/EPFL_TH8338.pdf?subformat=pdfa
and
http://www.hpcadvisorycouncil.com/events/2018/swiss-workshop/agenda.php
Sign up for our insideHPC Newsletter: http://insidehpc.com/newsletter
This document provides an overview and outline of a talk on quantum computing in practice and applications to cryptography. The talk will introduce quantum physics basics, discuss the state of quantum computing and cryptography, explain how to build quantum circuits, and provide tools and access for practicing quantum computing. It will cover fundamental quantum algorithms, attacks against cryptography, simulations and tools for quantum computing, and the future of post-quantum cryptography.
Regression and Classification: An Artificial Neural Network ApproachKhulna University
This presentation introduces artificial neural networks (ANN) as a technique for regression and classification problems. It provides historical context on the development of ANN, describes common network structures and activation functions, and the backpropagation algorithm for training networks. Experimental results on 7 datasets show ANN outperformed other methods for both regression and classification across a variety of problem types and data characteristics. Limitations of ANN and areas for further research are also discussed.
In this video from PASC18, Alexander Nitz from the Max Planck Institute for Gravitational Physics in Germany presents: The Search for Gravitational Waves.
"The LIGO and Virgo detectors have completed a prolific observation run. We are now observing gravitational waves from both the mergers of binary black holes and neutron stars. We’ll discuss how these discoveries were made and look into what the near future of searching for gravitational waves from compact binary mergers will look like."
Watch the video: https://wp.me/p3RLHQ-iTv
Learn more: github.com/gwastro/pycbc
and
https://pasc18.pasc-conference.org/
Sign up for our insideHPC Newsletter: http://insidehpc.com/newsletter
The document presents SURF (Speeded Up Robust Features), a novel scale- and rotation-invariant detector and descriptor. SURF uses a fast-Hessian detector based on the Hessian matrix and DoG approximations. It assigns orientations based on Haar wavelet responses and extracts 64-dimensional descriptors from summed Haar wavelet responses. SURF matches features based on sign of the Laplacian for fast indexing. Experiments show SURF outperforms other methods in repeatability, distinctiveness and robustness while computing faster.
Computing Just What You Need: Online Data Analysis and Reduction at Extreme ...Ian Foster
This document discusses computing challenges posed by rapidly increasing data scales in scientific applications and high performance computing. It introduces the concept of online data analysis and reduction as an alternative to traditional offline analysis to help address these challenges. The key messages are that dramatic changes in HPC system geography due to different growth rates of technologies are driving new application structures and computational logistics problems, presenting exciting new computer science opportunities in online data analysis and reduction.
The document discusses ensemble clustering methods. It begins by comparing classification and clustering, noting that clustering differs in that ground truth labels are not known beforehand. It then discusses how ensemble clustering can improve upon single clustering algorithms by generating multiple partitions and combining them. The key steps are: 1) generating an ensemble of initial partitions from clustering the data multiple times, 2) aligning the initial partitions into metaclusters, and 3) voting to determine a final clustering assignment. This approach provides benefits of scalability and robustness over single clustering algorithms.
Performance Optimization of CGYRO for Multiscale Turbulence SimulationsIgor Sfiligoi
Overview of the recent performance optimization of CGYRO, an Eulerian GyroKinetic Fusion Plasma solver, with emphasize on the Multiscale Turbulence Simulations.
Presented at the joint US-Japan Workshop on Exascale Computing Collaboration and6th workshop of US-Japan Joint Institute for Fusion Theory (JIFT) program (Jan 18th 2022).
https://telecombcn-dl.github.io/2018-dlcv/
Deep learning technologies are at the core of the current revolution in artificial intelligence for multimedia data analysis. The convergence of large-scale annotated datasets and affordable GPU hardware has allowed the training of neural networks for data analysis tasks which were previously addressed with hand-crafted features. Architectures such as convolutional neural networks, recurrent neural networks and Q-nets for reinforcement learning have shaped a brand new scenario in signal processing. This course will cover the basic principles and applications of deep learning to computer vision problems, such as image classification, object detection or image captioning.
The document discusses clustering techniques and provides details about the k-means clustering algorithm. It begins with an introduction to clustering and lists different clustering techniques. It then describes the k-means algorithm in detail, including how it works, the steps involved, and provides an example illustration. Finally, it discusses comments on the k-means algorithm, focusing on aspects like choosing the value of k, initializing cluster centroids, and different distance measurement methods.
Inside LoLA - Experiences from building a state space tool for place transiti...Universität Rostock
LoLA is a state space tool for analyzing place/transition nets that was developed starting in 1998. It uses various reduction techniques like stubborn sets, symmetries, and linear algebra to combat state space explosion. LoLA has been applied to problems in areas like model checking, business process verification, and distributed systems. Its core data structures and algorithms keep processing costs low during operations like firing transitions and state space traversal.
The document discusses hyperspectral imaging and architectural trades for a hyperspectral imaging system. It describes how hyperspectral data is collected and formatted. It discusses potential architectural approaches including spatial scanning, spectral binning, and a SIMD processor array. It provides block diagrams of example hyperspectral imaging systems and payloads. It also discusses requirements and implementation considerations for mapping hyperspectral data to Landsat equivalent bands.
Similar to LHCb Computing Workshop 2018: PV finding with CNNs (20)
Modern binary build systems have made shipping binary packages for Python much easier than ever before. This talk discusses three of the most popular build systems for Python packages using the new standards developed for packaging.
This document discusses software quality assurance tooling, focusing on pre-commit. It introduces pre-commit as a tool for running code quality checks before code is committed. Pre-commit allows configuring hooks that run checks and fixers on files matching certain patterns. Hooks can be installed from repositories and support many languages including Python. The document provides examples of pre-commit checks such as disallowing improper capitalization in code comments and files. It also discusses how to configure, run, update and install pre-commit hooks.
The document summarizes Henry Schreiner's work on several Python and C++ scientific computing projects. It describes a scientific Python development guide built from the Scikit-HEP summit. It also outlines Henry's work on pybind11 for C++ bindings, scikit-build for building extensions, cibuildwheel for building wheels on CI, and several other related projects.
Flake8 is a Python linter that is fast, simple, and extensible. It can be configured through setup.cfg or .flake8 files to ignore certain checks or select others. The summary recommends using the flake8-bugbear plugin and avoiding all print statements with flake8-print. Linters like Flake8 help find errors, improve code quality, and avoid historical baggage, but one does not need every check and it is okay to build a long ignore list.
The document describes various productivity tools for Python development, including:
- Pre-commit hooks to run checks before committing code
- Hot code reloading in Jupyter notebooks using the %load_ext and %autoreload magic commands
- Cookiecutter for generating project templates
- SSH configuration files and escape sequences for easier remote access
- Autojump to quickly navigate frequently visited directories
- Terminal tips like command history search and referencing the last argument
- Options for tracking Jupyter notebooks with git like stripping outputs or synchronizing notebooks and Python files.
SciPy22 - Building binary extensions with pybind11, scikit build, and cibuild...Henry Schreiner
Building binary extensions is easier than ever thanks to several key libraries. Pybind11 provides a natural C++ language for extensions without requiring pre-processing or special dependencies. Scikit-build ties the premier C++ build system, CMake, into the Python extension build process. And cibuildwheel makes it easy to build highly compatible wheels for over 80 different platforms using CI or on your local machine. We will look at advancements to all three libraries over the last year, as well as future plans.
This document discusses the history and development of Python packages for high energy physics (HEP) analysis. It describes how experiments initially used ROOT and C++, but Python gained popularity for configuration and analysis. This led to the creation of packages like Scikit-HEP, Uproot, and Awkward Array to bridge the gap between ROOT files and the Python data science stack. Scikit-HEP grew to include many related packages and provides best practices through its developer pages. The future may include adopting Scikit-build for building Python packages with C/C++ extensions and running packages in the browser via WebAssembly.
PyCon 2022 -Scikit-HEP Developer Pages: Guidelines for modern packagingHenry Schreiner
This was a PyCon 2022 lightning talk over the Scikit-HEP developer pages. It highlights best practices and guides shown there, and the quick package creation cookiecutter. And finally it demos the Pyodide WebAssembly app embedded into the Scikit-HEP developer pages!
Talk at PyCon2022 over building binary packages for Python. Covers an overview and an in-depth look into pybind11 for binding, scikit-build for creating the build, and build & cibuildwheel for making the binaries that can be distributed on PyPI.
Digital RSE: automated code quality checks - RSE group meetingHenry Schreiner
Given at a local RSE group meeting. Covers code quality practices, focusing on Python but over multiple languages, with useful tools highlighted throughout.
This document provides best practices for using CMake, including:
- Set the cmake_minimum_required version to ensure modern features while maintaining backward compatibility.
- Use targets to define executables and libraries, their properties, and dependencies.
- Fetch remote dependencies at configure time using FetchContent or integrate with package managers like Conan.
- Import library targets rather than reimplementing Find modules when possible.
- Treat CUDA as a first-class language in CMake projects.
HOW 2019: A complete reproducible ROOT environment in under 5 minutesHenry Schreiner
The document discusses setting up a ROOT environment using Conda in under 5 minutes. It describes downloading and installing Miniconda and then using Conda commands to create a new environment and install ROOT and its dependencies from the conda-forge channel. The ROOT package provides full ROOT functionality, including compilation and graphics, and supports Linux, macOS, and multiple Python versions.
2019 IRIS-HEP AS workshop: Boost-histogram and histHenry Schreiner
The document discusses the current state of histograms in Python and the need for a new histogramming library. It introduces boost-histogram, a C++ histogramming library, and its new Python bindings. The bindings aim to provide a fast, flexible and easily distributable histogram object for Python. Key features discussed include histogram design that treats it as a first-class object, fast filling via multi-threading, a variety of axis and storage types, and performance benchmarks showing it can be over 10x faster than NumPy for filling histograms. Distribution is focused on providing binary wheels for many platforms via continuous integration.
The document discusses the current state of histograms in Python and the need for a new library. It introduces boost-histogram, a C++ histogram library, and its new Python bindings. The bindings aim to provide a fast, flexible, and easily distributable histogram object for Python with support for multiple axis types and storage options. It also discusses plans for an additional wrapper library called hist for easy plotting and interfacing with other tools.
2019 IRIS-HEP AS workshop: Particles and decaysHenry Schreiner
The Scikit-HEP project aims to create an ecosystem for particle physics data analysis in Python. It includes packages like Particle and DecayLanguage that provide tools for working with particle data and decay descriptions. Particle allows users to easily access and search particle property data from sources like the PDG. DecayLanguage allows parsing decay file formats, representing and manipulating decay chains, and converting between decay model representations. Future work includes expanding particle ID support and improving visualization of decay trees.
Or: Beyond linear.
Abstract: Equivariant neural networks are neural networks that incorporate symmetries. The nonlinear activation functions in these networks result in interesting nonlinear equivariant maps between simple representations, and motivate the key player of this talk: piecewise linear representation theory.
Disclaimer: No one is perfect, so please mind that there might be mistakes and typos.
dtubbenhauer@gmail.com
Corrected slides: dtubbenhauer.com/talks.html
The technology uses reclaimed CO₂ as the dyeing medium in a closed loop process. When pressurized, CO₂ becomes supercritical (SC-CO₂). In this state CO₂ has a very high solvent power, allowing the dye to dissolve easily.
ESPP presentation to EU Waste Water Network, 4th June 2024 “EU policies driving nutrient removal and recycling
and the revised UWWTD (Urban Waste Water Treatment Directive)”
The cost of acquiring information by natural selectionCarl Bergstrom
This is a short talk that I gave at the Banff International Research Station workshop on Modeling and Theory in Population Biology. The idea is to try to understand how the burden of natural selection relates to the amount of information that selection puts into the genome.
It's based on the first part of this research paper:
The cost of information acquisition by natural selection
Ryan Seamus McGee, Olivia Kosterlitz, Artem Kaznatcheev, Benjamin Kerr, Carl T. Bergstrom
bioRxiv 2022.07.02.498577; doi: https://doi.org/10.1101/2022.07.02.498577
ESR spectroscopy in liquid food and beverages.pptxPRIYANKA PATEL
With increasing population, people need to rely on packaged food stuffs. Packaging of food materials requires the preservation of food. There are various methods for the treatment of food to preserve them and irradiation treatment of food is one of them. It is the most common and the most harmless method for the food preservation as it does not alter the necessary micronutrients of food materials. Although irradiated food doesn’t cause any harm to the human health but still the quality assessment of food is required to provide consumers with necessary information about the food. ESR spectroscopy is the most sophisticated way to investigate the quality of the food and the free radicals induced during the processing of the food. ESR spin trapping technique is useful for the detection of highly unstable radicals in the food. The antioxidant capability of liquid food and beverages in mainly performed by spin trapping technique.
Remote Sensing and Computational, Evolutionary, Supercomputing, and Intellige...University of Maribor
Slides from talk:
Aleš Zamuda: Remote Sensing and Computational, Evolutionary, Supercomputing, and Intelligent Systems.
11th International Conference on Electrical, Electronics and Computer Engineering (IcETRAN), Niš, 3-6 June 2024
Inter-Society Networking Panel GRSS/MTT-S/CIS Panel Session: Promoting Connection and Cooperation
https://www.etran.rs/2024/en/home-english/
PPT on Direct Seeded Rice presented at the three-day 'Training and Validation Workshop on Modules of Climate Smart Agriculture (CSA) Technologies in South Asia' workshop on April 22, 2024.
Current Ms word generated power point presentation covers major details about the micronuclei test. It's significance and assays to conduct it. It is used to detect the micronuclei formation inside the cells of nearly every multicellular organism. It's formation takes place during chromosomal sepration at metaphase.
Authoring a personal GPT for your research and practice: How we created the Q...Leonel Morgado
Thematic analysis in qualitative research is a time-consuming and systematic task, typically done using teams. Team members must ground their activities on common understandings of the major concepts underlying the thematic analysis, and define criteria for its development. However, conceptual misunderstandings, equivocations, and lack of adherence to criteria are challenges to the quality and speed of this process. Given the distributed and uncertain nature of this process, we wondered if the tasks in thematic analysis could be supported by readily available artificial intelligence chatbots. Our early efforts point to potential benefits: not just saving time in the coding process but better adherence to criteria and grounding, by increasing triangulation between humans and artificial intelligence. This tutorial will provide a description and demonstration of the process we followed, as two academic researchers, to develop a custom ChatGPT to assist with qualitative coding in the thematic data analysis process of immersive learning accounts in a survey of the academic literature: QUAL-E Immersive Learning Thematic Analysis Helper. In the hands-on time, participants will try out QUAL-E and develop their ideas for their own qualitative coding ChatGPT. Participants that have the paid ChatGPT Plus subscription can create a draft of their assistants. The organizers will provide course materials and slide deck that participants will be able to utilize to continue development of their custom GPT. The paid subscription to ChatGPT Plus is not required to participate in this workshop, just for trying out personal GPTs during it.
EWOCS-I: The catalog of X-ray sources in Westerlund 1 from the Extended Weste...Sérgio Sacani
Context. With a mass exceeding several 104 M⊙ and a rich and dense population of massive stars, supermassive young star clusters
represent the most massive star-forming environment that is dominated by the feedback from massive stars and gravitational interactions
among stars.
Aims. In this paper we present the Extended Westerlund 1 and 2 Open Clusters Survey (EWOCS) project, which aims to investigate
the influence of the starburst environment on the formation of stars and planets, and on the evolution of both low and high mass stars.
The primary targets of this project are Westerlund 1 and 2, the closest supermassive star clusters to the Sun.
Methods. The project is based primarily on recent observations conducted with the Chandra and JWST observatories. Specifically,
the Chandra survey of Westerlund 1 consists of 36 new ACIS-I observations, nearly co-pointed, for a total exposure time of 1 Msec.
Additionally, we included 8 archival Chandra/ACIS-S observations. This paper presents the resulting catalog of X-ray sources within
and around Westerlund 1. Sources were detected by combining various existing methods, and photon extraction and source validation
were carried out using the ACIS-Extract software.
Results. The EWOCS X-ray catalog comprises 5963 validated sources out of the 9420 initially provided to ACIS-Extract, reaching a
photon flux threshold of approximately 2 × 10−8 photons cm−2
s
−1
. The X-ray sources exhibit a highly concentrated spatial distribution,
with 1075 sources located within the central 1 arcmin. We have successfully detected X-ray emissions from 126 out of the 166 known
massive stars of the cluster, and we have collected over 71 000 photons from the magnetar CXO J164710.20-455217.
The use of Nauplii and metanauplii artemia in aquaculture (brine shrimp).pptxMAGOTI ERNEST
Although Artemia has been known to man for centuries, its use as a food for the culture of larval organisms apparently began only in the 1930s, when several investigators found that it made an excellent food for newly hatched fish larvae (Litvinenko et al., 2023). As aquaculture developed in the 1960s and ‘70s, the use of Artemia also became more widespread, due both to its convenience and to its nutritional value for larval organisms (Arenas-Pardo et al., 2024). The fact that Artemia dormant cysts can be stored for long periods in cans, and then used as an off-the-shelf food requiring only 24 h of incubation makes them the most convenient, least labor-intensive, live food available for aquaculture (Sorgeloos & Roubach, 2021). The nutritional value of Artemia, especially for marine organisms, is not constant, but varies both geographically and temporally. During the last decade, however, both the causes of Artemia nutritional variability and methods to improve poorquality Artemia have been identified (Loufi et al., 2024).
Brine shrimp (Artemia spp.) are used in marine aquaculture worldwide. Annually, more than 2,000 metric tons of dry cysts are used for cultivation of fish, crustacean, and shellfish larva. Brine shrimp are important to aquaculture because newly hatched brine shrimp nauplii (larvae) provide a food source for many fish fry (Mozanzadeh et al., 2021). Culture and harvesting of brine shrimp eggs represents another aspect of the aquaculture industry. Nauplii and metanauplii of Artemia, commonly known as brine shrimp, play a crucial role in aquaculture due to their nutritional value and suitability as live feed for many aquatic species, particularly in larval stages (Sorgeloos & Roubach, 2021).
The use of Nauplii and metanauplii artemia in aquaculture (brine shrimp).pptx
LHCb Computing Workshop 2018: PV finding with CNNs
1. PV finding with CNNs
LHCb Computing Workshop 2018
Rui Fang Henry Schreiner Mike Sokoloff
September 26, 2018
The University of Cincinnati
2. Objectives Introduction
Physics
• Iterative tracking and vertexing may allow
high efficiency, high speed, highly parallel
algorithms:
Use proto-tracks to find primary vertex
(PV) candidates
Use PV candidates to augment more
complete tracking
Find more PVs plus secondary vertices
• PVs available quickly
Machine learning
• Sparse 3D data (41M pixels) → rich 1D
dataset
• 1D convolutional neural net
• Great opportunities to visualize learning
process
Computation
• Highly parallelizable
• Well suited to GPUs
1/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
3. Tracking in the LHCb Upgrade Introduction
The changes
• 30 MHz software trigger
• 7.6 PVs per event (Poisson distribution)
The problem
• Much higher pileup
• Very little time to do the tracking
• Current algorithms too slow
We need to rethink our algorithms from the ground up...
2/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
4. A Hybrid ML Approach Introduction
Prototracking → Kernel generation → CNN to find PVs → Informed tracking
Prototracking
• Ultra-simple/fast
• Triplets only
• Used for kernel only
Vertexing
• High efficiency
• Low false positive rate
• Useful for other reasons
Tracking
• Faster (effect TBD)
• Uses search windows
• Higher efficiency
Machine learning features (so far)
• Prototracking converts sparse 3D dataset to feature-rich 1D dataset
• Easy and effective visualization due to 1D nature
• Can see results with simple unoptimized 2-layer CNN + 1-layer linear
What follows is a proof of principle implementation for finding PVs.
3/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
5. Vertices and Tracks Introduction
Vertices
• Events contain ≈ 7 Primary Vertices
(PVs)
A PV should contain 5+ long tracks
• Multiple Secondary Vertices (SVs) per
event as well
A SV should contain 2+ tracks
Beams
PV
Track
SV
• We are developing a way to find PVs and SVs using hit triplets
• This will enable an iterative tracking algorithm
4/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
6. Kernel Generation Design
Hits
• Hits lie on the 26 planes
• Tracks come from PVs and SVs
• For simplicity, only 3 tracks shown
• Hits are sorted in r (distance from LHC beam)
z axis (along the beam)
x PV
5/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
7. Kernel Generation Design
Grid
• Make a 3D grid of voxels (2D shown)
• Note: only z will be fully calculated and stored
z axis (along the beam)
x PV
5/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
8. Kernel Generation Design
Prototrack
• Start with maximum r
• Find triplet with χ2
< 10
• Mark “used” all other hits within χ2
< 9
• Note: triplet is stored
Kernel
• Fill in each voxel center with gaussian PDF
• PDF is combined for each prototrack
z axis (along the beam)
x PV
5/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
9. Kernel Generation Design
Prototrack
• Start with maximum r
• Find triplet with χ2
< 10
• Mark “used” all other hits within χ2
< 9
• Note: triplet is stored
Kernel
• Fill in each voxel center with gaussian PDF
• PDF is combined for each prototrack
z axis (along the beam)
x PV
5/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
10. Kernel Generation Design
Prototrack
• Start with maximum r
• Find triplet with χ2
< 10
• Mark “used” all other hits within χ2
< 9
• Note: triplet is stored
Kernel
• Fill in each voxel center with gaussian PDF
• PDF is combined for each prototrack
z axis (along the beam)
x PV
5/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
11. Kernel Generation Design
Kernel
• Highest PDF density at vertices
• Stores z histogram with maximum PDF values
Details
• x-y grid initially very coarse
• Search performed on maximum x-y grid cell
using stored triplets to recalculate PDF
z axis (along the beam)
x PV
5/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
12. Example of z KDE histogram Design
100 50 0 50 100 150 200 250 300
z values [mm]
0
500
1000
1500
2000
DensityofKernel
Kernel
LHCb PVs
Other PVs
LHCb SVs
Other SVs
Human learning
• Peaks generally correspond to PVs and SVs
Challenges
• Vertex may be offset from peak
• Vertices interact
6/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
13. Target distribution Design
Build target distribution
• real PV position as the mean of Gaussian
distribution
• σ(standard deviation) is 100 µm
• calculate the cdf of each bin around of the
mean, within ± 3 bins (± 300 µm )
7/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
14. Neural network architecture with two convolutional layers Design
• Activation function for hidden layers: Leaky ReLu
• Activation function for output layer: Sigmoid
8/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
15. Activation function Design
Activation function for hidden layers Activation function for output layer
9/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
16. Cost Function Design
Approach
• Cost function should be similar to
Cross-Entropy for y → 0, y → 1;
cost = − y ln ˆy + (1 − y) ln(1 − ˆy)
• Should be symmetric with respect to
r = (ˆy/y) & 1/r
Sum Over Bins
ri ≡ (ˆyi + )/(yi + ) (1)
zi ≡
2 ri
ri + 1/ri
(2)
our cost =
bins
− ln zi (3)
10/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
17. Cost, efficiency, and false positive rate: 2 convolutionial layers Results
11/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
18. Cost, efficiency, and false positive rate: 3 convolutionial layers Results
12/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
19. Compare Predictions with Targets (3 convolutional layers) Results
86.00 87.00 88.00 89.00 90.00
z values [mm]
0
50
100
150
200
250
300
350
400
KernelDensity
True: 88.062 mm
Event 0: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
99.00 100.00 101.00 102.00 103.00
z values [mm]
0
500
1000
1500
KernelDensity
True: 100.825 mm
Pred: 100.726 mm
: 99 µm
Event 0: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
135.00 136.00 137.00 138.00 139.00
z values [mm]
0
200
400
600
800
1000
1200
1400
1600
KernelDensity
True: 136.602 mm
Pred: 136.601 mm
: 2 µm
Event 0: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
182.00 183.00 184.00 185.00 186.00
z values [mm]
0
200
400
600
800
1000
1200
KernelDensity
True: 183.668 mm
Pred: 183.734 mm
: 66 µm
Event 0: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
13/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
20. Compare Predictions with Targets (3 convolutional layers) Results
60.00 61.00 62.00 63.00 64.00
z values [mm]
0
100
200
300
400
500
600
KernelDensity
Pred: 61.784 mm
Event 2: False positive
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
89.00 90.00 91.00 92.00 93.00
z values [mm]
0
200
400
600
800
1000
KernelDensity
True: 91.279 mm
Pred: 91.244 mm
: 35 µm
Event 2: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
109.00 110.00 111.00 112.00 113.00
z values [mm]
0
50
100
150
200
KernelDensity
True: 110.675 mm
Event 2: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
127.00 128.00 129.00 130.00 131.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 128.742 mm
Pred: 128.720 mm
: 22 µm
Event 2: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
14/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
21. Efficiencies and False Positive Rates Results
parameter 2 CVN Layers 3 CVN Layers 4 CVN Layers
(Efficiency) = TP
TP+FN
≈ 58% ≈ 70% ≈ 75%
False Positive rate = FP
number of events
≈ 0.07 ≈ 0.08 ≈ 0.13
Found Not found
Real PV True positive False negative
Not a real PV False positive True negative
True Positive
• search ±5 bins (±500µm) around a real PV
• at least 3 (4) bins with predicted probability > 1% and integrated probability > 20%.
False Positive
• at least 3 (4) bins with individual probabilities > 1% and integrated probability > 20%.
• no real PV within ±5 bins (±500µm) of that cluster.
15/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
22. Future Plans Future Plans
Some Ideas
• Model PV resolution as a function of the number of tracks; right now, the target function
is always generated assuming σz = 100 µm;
• Extract σz from predicted signals;
• Extend algorithm to find PV (x, y, z) target functions, not just PV z target functions.
• Mask PVs with < 5 long tracks (not labeled as PVs now)
• Ask the algorithm (very nicely) to find Secondary Vertices as well; it should probably use
both the original KDE histogram and the learned PV histogram as inputs. It may be
possible to re-use some of the features generated by the convolutional layers.
• Integrate KDE plus PV-finding code into an iterative tracking and vertexing algorithm;
well-defined vertex positions may be able to serve as anchors for good tracks, restricting
the roads to be searched.
• Optimize NN architecture to (i) improve learning, (ii) improve learning speed,
(iii) minimize inference costs (cycles and memory). Increase training sample.
16/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
23. Compare Predictions with Targets (3 convolutional layers) Backup
31.00 32.00 33.00 34.00 35.00
z values [mm]
0
200
400
600
800
KernelDensity
True: 32.975 mm
Pred: 32.766 mm
: 209 µm
Event 1: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
92.00 93.00 94.00 95.00 96.00
z values [mm]
0
50
100
150
200
250
300
350
400
KernelDensity
True: 93.727 mm
Event 1: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
107.00 108.00 109.00 110.00 111.00 112.00
z values [mm]
0
100
200
300
400
500
KernelDensity
True: 109.659 mm
Pred: 109.530 mm
: 129 µm
Event 1: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
116.00 117.00 118.00 119.00 120.00
z values [mm]
0
200
400
600
800
1000
KernelDensity
True: 118.176 mm
Pred: 118.224 mm
: 48 µm
Event 1: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Probability
Target
Predicted
17/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
24. Compare Predictions with Targets (3 convolutional layers) Backup
138.00 139.00 140.00 141.00 142.00
z values [mm]
0
500
1000
1500
KernelDensity
True: 139.889 mm
Pred: 139.861 mm
: 28 µm
Event 1: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Probability
Target
Predicted
152.00 153.00 154.00 155.00 156.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 153.906 mm
Pred: 153.864 mm
: 42 µm
Event 1: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
194.00 195.00 196.00 197.00 198.00
z values [mm]
0
50
100
150
200
250
KernelDensity
True: 195.838 mm
Event 1: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
222.00 223.00 224.00 225.00 226.00
z values [mm]
0
50
100
150
200
250
300
350
400
KernelDensity
True: 224.386 mm
Pred: 224.583 mm
: 197 µm
Event 1: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
18/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
25. More Predictions with Targets (3 CVN layers) Backup
159.00 160.00 161.00 162.00 163.00
z values [mm]
0
20
40
60
80
100
120
KernelDensity
True: 160.920 mm
Event 2: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
65.00 66.00 67.00 68.00 69.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 66.923 mm
Pred: 66.964 mm
: 42 µm
Event 3: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
66.00 67.00 68.00 69.00 70.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 67.930 mm
Event 3: PV not found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
108.00 109.00 110.00 111.00 112.00
z values [mm]
0
200
400
600
800
1000
1200
KernelDensity
True: 110.449 mm
Pred: 110.413 mm
: 36 µm
Event 3: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
19/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
26. More Predictions with Targets (3 CVN layers) Backup
120.00 121.00 122.00 123.00 124.00
z values [mm]
0
100
200
300
400
500
600
700
KernelDensity
True: 122.303 mm
Pred: 122.261 mm
: 42 µm
Event 3: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
208.00 209.00 210.00 211.00 212.00
z values [mm]
0
50
100
150
200
250
300
KernelDensity
True: 210.210 mm
Pred: 210.417 mm
: 207 µm
Event 3: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
239.00 240.00 241.00 242.00 243.00
z values [mm]
0
10
20
30
40
50
60
70
KernelDensity
True: 240.865 mm
Event 3: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
17.00 18.00 19.00 20.00 21.00
z values [mm]
0
50
100
150
KernelDensity
True: 19.440 mm
Event 4: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
20/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
27. More Predictions with Targets (3 CVN layers) Backup
71.00 72.00 73.00 74.00 75.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 72.816 mm
Pred: 72.764 mm
: 52 µm
Event 4: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
92.00 93.00 94.00 95.00 96.00
z values [mm]
0
100
200
300
400
KernelDensity
True: 93.940 mm
Pred: 93.959 mm
: 19 µm
Event 4: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
117.00 118.00 119.00 120.00 121.00
z values [mm]
0
200
400
600
800
KernelDensity
True: 119.325 mm
Pred: 119.513 mm
: 188 µm
Event 4: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
145.00 146.00 147.00 148.00 149.00
z values [mm]
0
50
100
150
200
250
300
350
400
KernelDensity
True: 146.840 mm
Pred: 146.923 mm
: 84 µm
Event 4: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
Probability
Target
Predicted
21/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
28. More Predictions with Targets (3 CVN layers) Backup
195.00 196.00 197.00 198.00 199.00
z values [mm]
0
50
100
150
200
250
300
350
400
KernelDensity
True: 197.279 mm
Pred: 197.330 mm
: 51 µm
Event 4: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
60.00 61.00 62.00 63.00 64.00
z values [mm]
0
100
200
300
400
500
600
700
800
KernelDensity
True: 62.436 mm
Pred: 62.428 mm
: 8 µm
Event 5: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
72.00 73.00 74.00 75.00 76.00
z values [mm]
0
500
1000
1500
2000
2500
KernelDensity
True: 73.592 mm
Pred: 73.594 mm
: 2 µm
Event 5: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
98.00 99.00 100.00 101.00 102.00
z values [mm]
0
500
1000
1500
2000
2500
KernelDensity
True: 100.143 mm
Pred: 100.129 mm
: 14 µm
Event 5: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
22/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
29. More Predictions with Targets (3 CVN layers) Backup
113.00 114.00 115.00 116.00 117.00
z values [mm]
0
200
400
600
800
1000
1200
KernelDensity
True: 114.670 mm
Pred: 114.756 mm
: 86 µm
Event 5: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
131.00 132.00 133.00 134.00 135.00
z values [mm]
0
50
100
150
200
250
KernelDensity
True: 132.683 mm
Pred: 132.622 mm
: 61 µm
Event 5: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
156.00 157.00 158.00 159.00 160.00
z values [mm]
0
500
1000
1500
KernelDensity
True: 158.211 mm
Pred: 158.333 mm
: 121 µm
Event 5: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
158.00 159.00 160.00 161.00 162.00
z values [mm]
0
500
1000
1500
KernelDensity
True: 159.703 mm
Event 5: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
23/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
30. More Predictions with Targets (3 CVN layers) Backup
176.00 177.00 178.00 179.00 180.00
z values [mm]
0
200
400
600
800
1000
1200
KernelDensity
True: 177.635 mm
Pred: 177.621 mm
: 14 µm
Event 5: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
10.00 11.00 12.00 13.00 14.00
z values [mm]
0
50
100
150
200
250
300
350
KernelDensity
True: 11.973 mm
Event 6: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
50.00 51.00 52.00 53.00 54.00 55.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 52.597 mm
Pred: 52.550 mm
: 47 µm
Event 6: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Probability
Target
Predicted
56.00 57.00 58.00 59.00 60.00 61.00
z values [mm]
0
100
200
300
400
500
600
700
KernelDensity
True: 58.490 mm
Pred: 58.494 mm
: 4 µm
Event 6: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
24/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
31. More Predictions with Targets (3 CVN layers) Backup
141.00 142.00 143.00 144.00 145.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 143.421 mm
Pred: 143.445 mm
: 24 µm
Event 6: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
147.00 148.00 149.00 150.00 151.00
z values [mm]
0
200
400
600
800
1000
1200
KernelDensity
True: 148.586 mm
Pred: 148.622 mm
: 36 µm
Event 6: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
229.00 230.00 231.00 232.00 233.00
z values [mm]
0
100
200
300
400
500
KernelDensity
True: 230.789 mm
Pred: 230.886 mm
: 97 µm
Event 6: PV found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
61.00 62.00 63.00 64.00 65.00
z values [mm]
0
50
100
150
200
KernelDensity
True: 62.706 mm
Event 7: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
25/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
32. More Predictions with Targets (3 CVN layers) Backup
101.00 102.00 103.00 104.00 105.00 106.00
z values [mm]
0
100
200
300
400
KernelDensity
True: 103.480 mm
Event 7: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
105.00 106.00 107.00 108.00 109.00
z values [mm]
0
200
400
600
800
1000
KernelDensity
True: 106.733 mm
Pred: 106.813 mm
: 80 µm
Event 7: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
165.00 166.00 167.00 168.00 169.00
z values [mm]
0
200
400
600
800
1000
KernelDensity
True: 167.395 mm
Pred: 167.434 mm
: 40 µm
Event 7: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
18.00 19.00 20.00 21.00 22.00
z values [mm]
0
500
1000
1500
2000
KernelDensity
True: 19.824 mm
Pred: 19.877 mm
: 53 µm
Event 8: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
26/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
33. More Predictions with Targets (3 CVN layers) Backup
113.00 114.00 115.00 116.00 117.00
z values [mm]
0
100
200
300
400
500
KernelDensity
True: 114.807 mm
Event 8: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
155.00 156.00 157.00 158.00 159.00
z values [mm]
0
200
400
600
800
1000
KernelDensity
True: 157.448 mm
Pred: 157.432 mm
: 16 µm
Event 8: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
161.00 162.00 163.00 164.00 165.00
z values [mm]
0
100
200
300
400
500
600
KernelDensity
True: 162.676 mm
Event 8: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
162.00 163.00 164.00 165.00 166.00
z values [mm]
0
100
200
300
400
500
600
KernelDensity
True: 163.772 mm
Event 8: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
27/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
34. More Predictions with Targets (3 CVN layers) Backup
228.00 229.00 230.00 231.00 232.00
z values [mm]
0
50
100
150
200
KernelDensity
True: 230.200 mm
Event 8: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
1.00 2.00 3.00 4.00 5.00
z values [mm]
0
500
1000
1500
2000
2500
3000
KernelDensity
True: 3.065 mm
Pred: 3.068 mm
: 3 µm
Event 9: PV found
Kernel Density
0.0
0.2
0.4
0.6
0.8
1.0
Probability
Target
Predicted
64.00 65.00 66.00 67.00 68.00
z values [mm]
0
50
100
150
200
250
300
350
KernelDensity
True: 65.737 mm
Event 9: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
153.00 154.00 155.00 156.00 157.00
z values [mm]
0
50
100
150
200
250
KernelDensity
True: 155.047 mm
Event 9: PV not found
Kernel Density
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Probability
Target
Predicted
28/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018
35. The VELO Backup
Tracks
• Originate from vertices (not shown)
• Hits originate from tracks
• We only know the true track in simulation
• Nearly straight, but tracks may scatter in material
The VELO
• A set of 26 planes that detect tracks
• Tracks should hit one or more pixels per plane
• Sparse 3D dataset (41M pixels)
29/16Fang, Schreiner, Sokoloff
PV finding with CNNs: LHCb Computing Workshop 2018
September 26, 2018