Machine Learning Developments in ROOT Sergei Gleyzer, Lorenzo Moneta

Slides:



Advertisements
Similar presentations
Neural networks Introduction Fitting neural networks
Advertisements

CSCI 347 / CS 4206: Data Mining Module 07: Implementations Topic 03: Linear Models.
Visualization of hidden node activity in a feed forward neural network Adam Arvay.
HPCC Mid-Morning Break High Performance Computing on a GPU cluster Dirk Colbry, Ph.D. Research Specialist Institute for Cyber Enabled Discovery.
CES 514 – Data Mining Lecture 8 classification (contd…)
Submitted by:Supervised by: Ankit Bhutani Prof. Amitabha Mukerjee (Y )Prof. K S Venkatesh.
1 1 © 2011 The MathWorks, Inc. Accelerating Bit Error Rate Simulation in MATLAB using Graphics Processors James Lebak Brian Fanous Nick Moore High-Performance.
Slide Image Retrieval: A Preliminary Study Guo Min Liew and Min-Yen Kan National University of Singapore Web IR / NLP Group (WING)
ROOT: A Data Mining Tool from CERN Arun Tripathi and Ravi Kumar 2008 CAS Ratemaking Seminar on Ratemaking 17 March 2008 Cambridge, Massachusetts.
Machine Learning1 Machine Learning: Summary Greg Grudic CSCI-4830.
Back-Propagation MLP Neural Network Optimizer ECE 539 Andrew Beckwith.
Michigan REU Final Presentations, August 10, 2006Matt Jachowski 1 Multivariate Analysis, TMVA, and Artificial Neural Networks Matt Jachowski
Comparison of Bayesian Neural Networks with TMVA classifiers Richa Sharma, Vipin Bhatnagar Panjab University, Chandigarh India-CMS March, 2009 Meeting,
CS 732: Advance Machine Learning
SFT Summer Student Proposals 9 th February CERN Summer Student Programme  PH has been allocated 91 MS students (SFT quota 6-7)  submit request.
Anders Nielsen Technical University of Denmark, DTU-Aqua Mark Maunder Inter-American Tropical Tuna Commission An Introduction.
Neural Network Analysis of Dimuon Data within CMS Shannon Massey University of Notre Dame Shannon Massey1.
TMVA New Features Sergei V. Gleyzer Omar Zapata Mesa Lorenzo Moneta Data Science at the LHC Workshop Nov. 9, 2015.
Frameworks and Collaborative Tools. The request (from Ruth’s ) … directions in the area of discussion of other scientific communities, open source.
Feature selection using Deep Neural Networks March 18, 2016 CSI 991 Kevin Ham.
Jun Doi IBM Research – Tokyo Early Performance Evaluation of Lattice QCD on POWER+GPU Cluster 17 July 2015.
9/24/2017 7:27 AM © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN.
IML News Sergei Gleyzer (CMS)
Comparing TensorFlow Deep Learning Performance Using CPUs, GPUs, Local PCs and Cloud Pace University, Research Day, May 5, 2017 John Lawrence, Jonas Malmsten,
Combining Models Foundations of Algorithms and Machine Learning (CS60020), IIT KGP, 2017: Indrajit Bhattacharya.
Effects of Limiting Numerical Precision on Neural Networks
TensorFlow The Deep Learning Library You Should Be Using.
TensorFlow CS 5665 F16 practicum Karun Joseph, A Reference:
Deep Learning Software: TensorFlow
World’s fastest Machine Learning Engine
Inter-experimental LHC Machine Learning Working Group Activities
Data Analytics and CERN IT Hadoop Service
Low-Cost High-Performance Computing Via Consumer GPUs
Hadoop and Analytics at CERN IT
Working With Azure Batch AI
Speaker Classification through Deep Learning
Sergei V. Gleyzer University of Florida
Deep Learning Insights and Open-ended Questions
Enabling machine learning in embedded systems
Deep Learning in HEP Large number of applications:
IML News IML WG Meeting Sergei Gleyzer (CMS)
Matt Gormley Lecture 16 October 24, 2016
Deep Learning Libraries
dawn.cs.stanford.edu/benchmark
Multi-Layer Perceptron On A GPU
Basic machine learning background with Python scikit-learn
Comparison Between Deep Learning Packages
Policy Compression for MDPs
Deep Learning Workshop
Intro to Machine Learning
Brewing Deep Networks With Caffe
Deep Learning Packages
Machine Learning with Weka
Introduction to Deep Learning with Keras
What's New in eCognition 9
CSC 578 Neural Networks and Deep Learning
Parallel k-means++ for Multiple Shared-Memory Architectures
Logistics Project proposals are due by midnight tonight (two pages)
Keras.
TensorFlow: A System for Large-Scale Machine Learning
Deep Learning Libraries
What's New in eCognition 9
What's New in eCognition 9
CSC 578 Neural Networks and Deep Learning
Single Parameter Tuning
LHC beam mode classification
Igor Stančin, Alan Jović to: {igor.stancin,
Jetson-Enabled Autonomous Vehicle
Advisor: Dr.vahidipour Zahra salimian Shaghayegh jalali Dec 2017
Machine Learning for Cyber
Presentation transcript:

Machine Learning Developments in ROOT Sergei Gleyzer, Lorenzo Moneta for the ROOT-TMVA Team CHEP 2016, October 10, 2016 Software Development

Outline Status and Overview New TMVA Features Future Plans and Outlook External Interfaces Deep Learning, Jupyter, Parallelization Future Plans and Outlook Summary 10/10/2016 Sergei V. Gleyzer CHEP 2016

TMVA Toolkit for Multivariate Analysis: HEP Machine Learning workhorse Part of ROOT In LHC experiments production Easy for beginners, powerful for experts 17 active contributors (5 GSoCs) 10/10/2016 Sergei V. Gleyzer CHEP 2016

New TMVA version released in upcoming ROOT 6.0.8 10/10/2016 Sergei V. Gleyzer CHEP 2016

New TMVA Features 10/10/2016 Sergei V. Gleyzer CHEP 2016

New Features Modularity, External Interfaces, Updated SVMs Analyzer Tools: Variable Importance Deep Learning CPU, GPU Parallelization with multithreading and GPUs Analyzer Tools: Cross-Validation, Hyper-Parameter Tuning Regression Loss Functions Jupyter: Interactive Training, Visualizations Unsupervised Learning Deep Autoencoders Multi-processing, Spark parallelization Added in 2015 Added in TMVA ROOT 6.0.8 Upcoming 2016 10/10/2016 Sergei V. Gleyzer CHEP 2016

TMVA Interfaces Interfaces to External ML Tools RMVA interface to R PyMVA interface to scikit-learn KMVA interface to Keras High-level interface to Theano, TensorFlow deep-learning libraries 10/10/2016 Sergei V. Gleyzer CHEP 2016

Deep Learning Is a powerful Machine Learning method based on Deep Neural Networks (DNN) that achieves significant performance improvement in classification tasks 10/10/2016 Sergei V. Gleyzer CHEP 2016

Deep Learning New Deep-Learning Library in TMVA GPU support CUDA OpenCL Excellent performance and high numerical throughput 10/10/2016 Sergei V. Gleyzer CHEP 2016

Deep Learning CPU Performance: Implementation: OpenBLAS, TBB Peak performance per core: 16 GFLOP/s Single, Double Precision 10/10/2016 Sergei V. Gleyzer CHEP 2016

Deep Learning GPU Performance: 20 input nodes Network: 20 input nodes 5 hidden layers with nh nodes each Hardware: NVIDIA Tesla K20 1.17 TFLOP/s peak performance @ double precision 2 compute streams - - - 1 compute stream Good Throughput 10/10/2016 Sergei V. Gleyzer CHEP 2016

Deep Learning Throughput Comparison 2.7 * Theano Single precision batch size = 1024 Excellent throughput compared to Theano on same GPU 10/10/2016 Sergei V. Gleyzer CHEP 2016

Deep Learning ROC Performance: significant improvements compared to shallow networks and boosted decision trees 10/10/2016 Sergei V. Gleyzer CHEP 2016

Cross Validation New features: k-fold cross-validation Hyper-parameter tuning Find optimized parameters (SVM, BDT) 10/10/2016 Sergei V. Gleyzer CHEP 2016

Regression New Regression Features: Loss functions: Huber (default) Least Squares Absolute Deviation Custom Function Important for regression performance Higher is better 10/10/2016 Sergei V. Gleyzer CHEP 2016

Jupyter Integration 10/10/2016 Sergei V. Gleyzer CHEP 2016

Pre-processing New pre-processing features: Hessian Locally Linear Embedding (Hessian LLE) Variance Threshold 10/10/2016 Sergei V. Gleyzer CHEP 2016

Some Upcoming Features 10/10/2016 Sergei V. Gleyzer CHEP 2016

Spark TMVA SPARK Parallelization Test Good speed-up in prototype R&D 10/10/2016 Sergei V. Gleyzer CHEP 2016

Deep Autoencoder 10/10/2016 Sergei V. Gleyzer CHEP 2016

Summary Many new features in TMVA release upcoming in ROOT 6.0.8 Production-ready parallelized Deep Learning Cross-validation, Hyper-parameter tuning Jupyter integration More pre-processing features Regression updates Many contributions Feedback and further contributions welcome 10/10/2016 Sergei V. Gleyzer CHEP 2016

Feature Contributors Sergei Gleyzer Analyzer Tools, Algorithm Development Lorenzo Moneta Multi-threading, Multi-processing Omar Zapata Mesa PyMVA, RMVA, Modularity, Parallelization Peter Speckmeyer Deep-Learning CPU Simon Pfreundschuh Deep-Learning CPU and GPU Adrian Bevan, Tom Stevenson SVMs, Cross-Validation, Hyperparameter Tuning Attila Bagoly Jupyter Integration, Visualization, Output Albulena Saliji TMVA Output Transformation Stefan Wunsch KERAS Interfance Pourya Vakilipourtakalou Cross-Validation, Parallelization Abhinav Moudhil Pre-processing, Deep Autoencoders Georgios Douzas Spark, Cross-Validation, Hyperparameter Tuning Paul Seyfert Performance optimization of MLP Andrew Carnes Regression, Loss Functions, BDT Parallelization Continued invaluable contributions from Andreas Hoecker, Helge Voss, Eckhard von Thorne, Jörg Stelzer, and key support from CERN EP-SFT Group 10/10/2016 Sergei V. Gleyzer CHEP 2016

More Information Websites: http://root.cern.ch http://iml.cern.ch http://oproject.org 10/10/2016 Sergei V. Gleyzer CHEP 2016

Inter-experimental LHC Machine Learning working group Exchange of HEP-ML expertise and experience among LHC experiments ML Forum ML software development and maintenance Exchange between HEP and ML communities Education (Tutorials) 02/09/16 Sergei V. Gleyzer QCHS 2016

Backup 10/10/2016 Sergei V. Gleyzer CHEP 2016

TMVA Deep Learning link 10/10/2016 Sergei V. Gleyzer CHEP 2016