DOCSLIB.ORG

Microsoft ML for Apache Spark

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Microsoft ML for Apache Spark Microsoft ML for Apache Spark Unifying Machine Learning Ecosystems at Massive Scales Mark Hamilton Microsoft, MIT [email protected] Overview Background Spark + SparkML MMLSpark Unifying ML Ecosystems LightGBM, CNTK, Vowpal Wabbit Vowpal Wabbit LightGBM CNTK Multilingual Bindings Microservice Orchestration Cognitive Services on Spark Cognitive Services Kubernetes Model Deployment with Spark Serving Use Cases The Snow Leopard Trust ● Code Driver ● Data Worker Worker Worker Age: Name: Age: Name: Age: Name: Int String Int String Int String A fault-tolerant distributed 15 Bob 1 Sam 25 Bob computing framework 25 Alice 2 Claire 25 Tom 33 Mary 53 Tina 66 Ted Map Reduce + SQL Data Source Whole program optimization + SQL, AWS Bucket, Local, CosmosDB, etc… query pushdown 120 110 Elastic 100 80 Scala, Python, R, Java, Julia 60 40 20 ML, Graph Processing, Streaming 0.9 0 Running Time(s) Hadoop Spark ML High level library for distributed machine learning More general than SciKit-Learn All models have a uniform interface Load Data Can compose models into Tokenizer Load Data complex pipelines Term Hashing Pipeline Can save, load, and transport Logistic Reg Evaluate models ● Estimator Evaluate ● Transformer Microsoft Machine Learning for Apache Spark v0.18 Microsoft’s Open Source Contributions to Apache Spark #UnifiedAnalytics #SparkAISummit 5 Distributed Fast Model Microservice Multilingual Binding Machine Learning Deployment Orchestration Generation www.aka.ms/spark Azure/mmlspark Unifying Machine Learning Ecosystems Goals Same API Composable Batch, Streaming, Serving Image Processing Gradient Boosting Deep Learning with Open CV with LightGBM Pipelines (Databricks) Elastically Distributed Fault Tolerant Multi-Language Data Source Agnostic Markus Cozowicz [email protected] Text Analytics with Distributed Model Deep Learning Data Scientist Vowpal Wabbit Interpretability with LIME with CNTK Example Backend: LightGBM on Spark PySpark SparklyR Wrapper Generation Barrier Execution for Synchronizing Workers Spark Worker Spark Worker LightGBM on Spark LGBM LGBM Spark Estimator/Transformer Fast Socket/MPI Process Process LightGBM Scala Bindings communication Scala/Java Interop mapPartitions for MPI Ring LightGBM Java Bindings Transformer LGBM SWIG Process Framework Time(s) Area under LightGBM Core (C++) ROC Spark Worker XGBoost 52.60 .808 SparkML GBT 82.78 .788 LightGBM 45.39 .812 Ilya Matiach, [email protected] Developer, Azure ML Cognitive Services Person High quality pre-built Skateboa intelligent services rd No time intensive model training or deployment “Hello World” Leverage Microsoft Research and Azure ML Place Time Range Available as Docker I had a wonderful trip to Seattle last week and even visited the Space Needle 2 times! Containers En-US 84% positive Place Vision Speech Language Decision Search Object, scene, and Speech transcription Language detection Q&A extraction from Ad-free web, news, image, activity detection (speech-to-text) unstructured text and video search results Named entity recognition Face recognition Custom speech models for Knowledge base creation Trends for video, news and identification unique vocabularies or Key phrase extraction from collections of Q&As complex environment Image identification, Celebrity and landmark Text sentiment analysis Semantic matching for classification and recognition Text-to-speech knowledge extraction Multilingual and contextual knowledge bases Emotion recognition Custom Voice spell checking Customizable content Identification of similar personalization learning images and products Text and handwriting Real-time speech translation Explicit or offensive text recognition (OCR) content moderation Named entity recognition Customizable speech and classification Customizable image transcription and translation PII detection for text recognition moderation Knowledge acquisition Speaker identification for named entities Video metadata, audio, and verification Text translation and keyframe extraction Search query autosuggest Customizable text translation and analysis Ad-free custom search Contextual language Explicit or offensive engine creation understanding content moderation Azure Cognitive Services on Spark Easy to use integration between Spark and the Azure Cognitive Services val df = new TextSentiment() Composable and pipelinable with all other SparkML models! .setTextCol(“text”) .setOutputCol(“sentiment”) Exponential Backoffs, .transform(inputs) Backpressure, Batching, Async Parallelism Features Time (s) Errors # None Fully Fluent API 30.8 18993 EBO+BP 1163.0 0 EBO+BP+B 57.1 0 EBO+BP+B+P 49.7 0 HTTP on Spark Web Service Full Integration between HTTP Protocol and Spark SQL Client Client Client Spark as a Microservice Partition Partition Partition Orchestrator Spark Worker Spark + X df = SimpleHTTPTransformer() .setInputParser(JSONInputParser()) Support for all Spark .setOutputParser(JSONOutputParser() Languages .setDataType(schema)) .setOutputCol("results") .setUrl(…) Deploying on Kubernetes Kubernetes (AKS, ACS, GKE, On-Prem etc) Works on any k8s cluster K8s worker K8s worker K8s worker Cognitive CognitiveSpark Cognitive Helm: Package Manager for Service ServiceWorker Service Cloud Container Container Container Cognitive Kubernetes Services HTTP on Spark HTTP on Spark HTTP on Spark HTTP on Spark Spark Spark Spark helm repo add microsoft \ Worker Worker Worker Storage or https://microsoft.github.io/charts/repo other Spark Serving Databases helm update Jupyter, Zepplin Spark Zepplin, Spark Readers Serving Load LIVY, or Balancer Spark Jupyter helm install microsoft/spark --version 1.0.0 Submit LB REST Requests to Submit Jobs, Run Notebooks, Deployed Models Manage Cluster, etc Users / Apps Dalitso Banda, [email protected] Microsoft AI Development Acceleration Program 12 Model Deployment with Spark Serving Sub-millisecond RESTful Model Deployment on Spark Clusters Users / Apps Batch API: Load Balancer HTTP Requests and spark.read.parquet.load(…) Responses .select(…) Server Server Streaming API: spark.readStream.kafka.load(…) Partition Partition Partition Partition Partition Partition .select(…) Spark Worker Spark Worker Serving API: Spark Master spark.readStream.server(“0.0.0.0”, 5000).load(…) .select(…) AI for Earth Endangered Status Matters Remote Camera Trapping Creating a labelled Training Dataset Creating a labelled Training Dataset Transfer Learning with ResNet 50 Low Level Mid Level Logistic Input Images Output Classes Features Features Regression Other Leopard Any SparkML Predictions Model Filters from Zeiler + Fergus 2013 Performance Without Deep With Deep Featurization, Featurization Augmentation, and Temporal Ensembling Accuracy 65.6% Accuracy 94.7% Goal: Identify Individual Leopards Source: HotSpotter - Patterned Species Instance Recognition Automating Detection with LIME on Spark LIME on Spark End to End Architecture Results Unsupervised Unsupervised Human Labels FRCNN Outputs Human Labels FRCNN Outputs Microsoft Machine Learning for Apache Spark v0.18 Microsoft’s Open Source Contributions to Apache Spark #UnifiedAnalytics #SparkAISummit 26 Distributed Fast Model Microservice Multilingual Binding Machine Learning Deployment Orchestration Generation www.aka.ms/spark Azure/mmlspark Thanks to Get in Touch You all! Ilya Matiach: LightGBM on Spark Support: mmlspark- Markus Cozowicz: VW on Spark [email protected] Sudarshan Raghunathan, Christina Lee, Daniel Ciborowski, Eli Barzilay, Tong Wen, Pablo Castro, Me: [email protected] Chris Hoder, Ryan Gaspar, Henrik Neilsen, Andrew Schonhoffer, Joseph Sirosh Github : Azure/mmlspark Microsoft NERD Garage Team + MIT Externship Website: www.aka.ms/spark Program Snow Leopard Trust: Koustubh Sharma, Rhetick Paper: www.aka.ms/spark-paper Sengupta, Jeff Brown, Michael Despines Microsoft Development Acceleration Team: Contributions Welcome! Dalitso Banda, Casey Hong, Karthik Rajendran, Check out our MSR Podcast on Manon Knoertzer, Tayo Amuneke, Alejandro Buendia Oct 2 Azure CAT, AzureML, and Azure Search Teams Backup Slides AI for Cultural Institutions Celebrating 2 years of Open Access at The MET In 2016 The MET Released 400k MIT, The MET, and Microsoft images under open access participated in a 3-day hackathon to create intelligent experiences using This past winter the MET released a the new collection new subject-keyword dataset of image annotations https://gen.studio 30 Goals: Create new Use new work Explore further works of art to explore with intelligent existing art search Needed Technologies: Generative Reverse image Elasticsearch Adversarial search with Cognitive Networks Services Generative Adversarial Art Training Data Real or Generated? Noise Real or Vector Generated? https://gen.studioGenerated Generator Discriminator Image Custom Reverse Image Search Query ResNet Deep Fast Nearest Closest Image Featurizer Features Neighbor Match Lookup MMLSpark SparkML LSH or Annoy Filters from Zeiler + Fergus 2013 Nearest Example Nearest Neighbors Example Nearest Neighbors Query Images Intelligent Search Index Pipe images through Query Computer Vision API Image: to annotate image for Describe searching A fish Image A picture A picture swimming Stream images and Output: containing a containing a intelligent person glass, cup underwater annotations to Azure Search Deep Feature Nearest Neighbors: End Application: Gen Studio https://gen.studio 36 AI for Accessibility Seeing AI Currency Identification A Familiar Architecture… Labelled Images Deep Features $1 $5 $2 Queries 0 1 Dollar Bing Azure Machine Union + Logistic 5 Dollars Image MobileNet Learning + Distinct Regression 10 Dollars Search Spark Serving 20 Dollars Prep Data Train Deploy.
Load more

Recommended publications
  • Asset Finder: a Search Tool for Finding Relevant Graphical Assets Using Automated Image Labelling
    Sjalvst¨ andigt¨ arbete i informationsteknologi 20 juni 2019 Asset Finder: A Search Tool for Finding Relevant Graphical Assets Using Automated Image Labelling Max Perea During¨ Anton Gildebrand Markus Linghede Civilingenjorsprogrammet¨ i informationsteknologi Master Programme in Computer and Information Engineering Abstract Institutionen for¨ Asset Finder: A Search Tool for Finding Rele- informationsteknologi vant Graphical Assets Using Automated Image Besoksadress:¨ Labelling ITC, Polacksbacken Lagerhyddsv¨ agen¨ 2 Max Perea During¨ Postadress: Box 337 Anton Gildebrand 751 05 Uppsala Markus Linghede Hemsida: http:/www.it.uu.se The creation of digital 3D-environments appears in a variety of contexts such as movie making, video game development, advertising, architec- ture, infrastructure planning and education. When creating these envi- ronments it is sometimes necessary to search for graphical assets in big digital libraries by trying different search terms. The goal of this project is to provide an alternative way to find graphical assets by creating a tool called Asset Finder that allows the user to search using images in- stead of words. The Asset Finder uses image labelling provided by the Google Vision API to find relevant search terms. The tool then uses synonyms and related words to increase the amount of search terms us- ing the WordNet database. Finally the results are presented in order of relevance using a score system. The tool is a web application with an in- terface that is easy to use. The results of this project show an application that is able to achieve good results in some of the test cases. Extern handledare: Teddy Bergsman Lind, Quixel AB Handledare: Mats Daniels, Bjorn¨ Victor Examinator: Bjorn¨ Victor Sammanfattning Skapande av 3D-miljoer¨ dyker upp i en mangd¨ olika kontexter som till exempel film- skapande, spelutveckling, reklam, arkitektur, infrastrukturplanering och utbildning.
    [Show full text]
  • A Tensor Compiler for Unified Machine Learning Prediction Serving
    A Tensor Compiler for Unified Machine Learning Prediction Serving Supun Nakandala, UC San Diego; Karla Saur, Microsoft; Gyeong-In Yu, Seoul National University; Konstantinos Karanasos, Carlo Curino, Markus Weimer, and Matteo Interlandi, Microsoft https://www.usenix.org/conference/osdi20/presentation/nakandala This paper is included in the Proceedings of the 14th USENIX Symposium on Operating Systems Design and Implementation November 4–6, 2020 978-1-939133-19-9 Open access to the Proceedings of the 14th USENIX Symposium on Operating Systems Design and Implementation is sponsored by USENIX A Tensor Compiler for Unified Machine Learning Prediction Serving Supun Nakandalac,∗, Karla Saurm, Gyeong-In Yus,∗, Konstantinos Karanasosm, Carlo Curinom, Markus Weimerm, Matteo Interlandim mMicrosoft, cUC San Diego, sSeoul National University {<name>.<surname>}@microsoft.com,[email protected], [email protected] Abstract TensorFlow [13] combined, and growing faster than both. Machine Learning (ML) adoption in the enterprise requires Acknowledging this trend, traditional ML capabilities have simpler and more efficient software infrastructure—the be- been recently added to DNN frameworks, such as the ONNX- spoke solutions typical in large web companies are simply ML [4] flavor in ONNX [25] and TensorFlow’s TFX [39]. untenable. Model scoring, the process of obtaining predic- When it comes to owning and operating ML solutions, en- tions from a trained model over new data, is a primary con- terprises differ from early adopters in their focus on long-term tributor to infrastructure complexity and cost as models are costs of ownership and amortized return on investments [68]. trained once but used many times. In this paper we propose As such, enterprises are highly sensitive to: (1) complexity, HUMMINGBIRD, a novel approach to model scoring, which (2) performance, and (3) overall operational efficiency of their compiles featurization operators and traditional ML models software infrastructure [14].
    [Show full text]
  • Naiad: a Timely Dataflow System
    Naiad: A Timely Dataflow System Derek G. Murray Frank McSherry Rebecca Isaacs Michael Isard Paul Barham Mart´ın Abadi Microsoft Research Silicon Valley {derekmur,mcsherry,risaacs,misard,pbar,abadi}@microsoft.com Abstract User queries Low-latency query are received responses are delivered Naiad is a distributed system for executing data parallel, cyclic dataflow programs. It offers the high throughput Queries are of batch processors, the low latency of stream proces- joined with sors, and the ability to perform iterative and incremental processed data computations. Although existing systems offer some of Complex processing these features, applications that require all three have re- incrementally re- lied on multiple platforms, at the expense of efficiency, Updates to executes to reflect maintainability, and simplicity. Naiad resolves the com- data arrive changed data plexities of combining these features in one framework. A new computational model, timely dataflow, under- Figure 1: A Naiad application that supports real- lies Naiad and captures opportunities for parallelism time queries on continually updated data. The across a wide class of algorithms. This model enriches dashed rectangle represents iterative processing that dataflow computation with timestamps that represent incrementally updates as new data arrive. logical points in the computation and provide the basis for an efficient, lightweight coordination mechanism. requirements: the application performs iterative process- We show that many powerful high-level programming ing on a real-time data stream, and supports interac- models can be built on Naiad’s low-level primitives, en- tive queries on a fresh, consistent view of the results. abling such diverse tasks as streaming data analysis, it- However, no existing system satisfies all three require- erative machine learning, and interactive graph mining.
    [Show full text]
  • Search, Analyse, Predict Image Spread on Twitter
    Image Search for Improved Law and Order: Search, Analyse, Predict image spread on Twitter Student Name: Sonal Goel IIIT-D-MTech-CS-GEN-16-MT14026 April, 2016 Indraprastha Institute of Information Technology New Delhi Thesis Committee Dr. Ponnurangam Kumaraguru (Chair) Dr. AV Subramanyam Dr. Samarth Bharadwaj Submitted in partial fulfillment of the requirements for the Degree of M.Tech. in Computer Science, in General Category ©2016 IIIT-D-MTech-CS-GEN-16-MT14026 All rights reserved Keywords: Image Search, Image Virality, Twitter, Law and Order, Police Certificate This is to certify that the thesis titled “Image Search for Improved Law and Order:Search, Analyse, Predict Image Spread on Twitter" submitted by Sonal Goel for the partial ful- fillment of the requirements for the degree of Master of Technology in Computer Science & Engineering is a record of the bonafide work carried out by her under our guidance and super- vision in the Security and Privacy group at Indraprastha Institute of Information Technology, Delhi. This work has not been submitted anywhere else for the reward of any other degree. Dr. Ponnurangam Kumarguru Indraprastha Institute of Information Technology, New Delhi Abstract Social media is often used to spread images that can instigate anger among people, hurt their religious, political, caste, and other sentiments, which in turn can create law and order situation in society. This results in the need for Law Enforcement Agencies (LEA) to inspect the spread of images related to such events on social media in real time. To help LEA analyse the image spread on microblogging websites like Twitter, we developed an Open Source Real Time Image Search System, where the user can give an image, and a supportive text related to image and the system finds the images that are similar to the input image along with their occurrences.
    [Show full text]
  • Lightgbm Release 3.2.1.99
    LightGBM Release 3.2.1.99 Microsoft Corporation Sep 26, 2021 CONTENTS: 1 Installation Guide 3 2 Quick Start 21 3 Python-package Introduction 23 4 Features 29 5 Experiments 37 6 Parameters 43 7 Parameters Tuning 65 8 C API 71 9 Python API 99 10 Distributed Learning Guide 175 11 LightGBM GPU Tutorial 185 12 Advanced Topics 189 13 LightGBM FAQ 191 14 Development Guide 199 15 GPU Tuning Guide and Performance Comparison 201 16 GPU SDK Correspondence and Device Targeting Table 205 17 GPU Windows Compilation 209 18 Recommendations When Using gcc 229 19 Documentation 231 20 Indices and Tables 233 Index 235 i ii LightGBM, Release 3.2.1.99 LightGBM is a gradient boosting framework that uses tree based learning algorithms. It is designed to be distributed and efficient with the following advantages: • Faster training speed and higher efficiency. • Lower memory usage. • Better accuracy. • Support of parallel, distributed, and GPU learning. • Capable of handling large-scale data. For more details, please refer to Features. CONTENTS: 1 LightGBM, Release 3.2.1.99 2 CONTENTS: CHAPTER ONE INSTALLATION GUIDE This is a guide for building the LightGBM Command Line Interface (CLI). If you want to build the Python-package or R-package please refer to Python-package and R-package folders respectively. All instructions below are aimed at compiling the 64-bit version of LightGBM. It is worth compiling the 32-bit version only in very rare special cases involving environmental limitations. The 32-bit version is slow and untested, so use it at your own risk and don’t forget to adjust some of the commands below when installing.
    [Show full text]
  • ML Cheatsheet Documentation
    ML Cheatsheet Documentation Team Sep 02, 2021 Basics 1 Linear Regression 3 2 Gradient Descent 21 3 Logistic Regression 25 4 Glossary 39 5 Calculus 45 6 Linear Algebra 57 7 Probability 67 8 Statistics 69 9 Notation 71 10 Concepts 75 11 Forwardpropagation 81 12 Backpropagation 91 13 Activation Functions 97 14 Layers 105 15 Loss Functions 117 16 Optimizers 121 17 Regularization 127 18 Architectures 137 19 Classification Algorithms 151 20 Clustering Algorithms 157 i 21 Regression Algorithms 159 22 Reinforcement Learning 161 23 Datasets 165 24 Libraries 181 25 Papers 211 26 Other Content 217 27 Contribute 223 ii ML Cheatsheet Documentation Brief visual explanations of machine learning concepts with diagrams, code examples and links to resources for learning more. Warning: This document is under early stage development. If you find errors, please raise an issue or contribute a better definition! Basics 1 ML Cheatsheet Documentation 2 Basics CHAPTER 1 Linear Regression • Introduction • Simple regression – Making predictions – Cost function – Gradient descent – Training – Model evaluation – Summary • Multivariable regression – Growing complexity – Normalization – Making predictions – Initialize weights – Cost function – Gradient descent – Simplifying with matrices – Bias term – Model evaluation 3 ML Cheatsheet Documentation 1.1 Introduction Linear Regression is a supervised machine learning algorithm where the predicted output is continuous and has a constant slope. It’s used to predict values within a continuous range, (e.g. sales, price) rather than trying to classify them into categories (e.g. cat, dog). There are two main types: Simple regression Simple linear regression uses traditional slope-intercept form, where m and b are the variables our algorithm will try to “learn” to produce the most accurate predictions.
    [Show full text]
  • 1.2 Le Zhang, Microsoft
    Objective • “Taking recommendation technology to the masses” • Helping researchers and developers to quickly select, prototype, demonstrate, and productionize a recommender system • Accelerating enterprise-grade development and deployment of a recommender system into production • Key takeaways of the talk • Systematic overview of the recommendation technology from a pragmatic perspective • Best practices (with example codes) in developing recommender systems • State-of-the-art academic research in recommendation algorithms Outline • Recommendation system in modern business (10min) • Recommendation algorithms and implementations (20min) • End to end example of building a scalable recommender (10min) • Q & A (5min) Recommendation system in modern business “35% of what consumers purchase on Amazon and 75% of what they watch on Netflix come from recommendations algorithms” McKinsey & Co Challenges Limited resource Fragmented solutions Fast-growing area New algorithms sprout There is limited reference every day – not many Packages/tools/modules off- and guidance to build a people have such the-shelf are very recommender system on expertise to implement fragmented, not scalable, scale to support and deploy a and not well compatible with enterprise-grade recommender by using each other scenarios the state-of-the-arts algorithms Microsoft/Recommenders • Microsoft/Recommenders • Collaborative development efforts of Microsoft Cloud & AI data scientists, Microsoft Research researchers, academia researchers, etc. • Github url: https://github.com/Microsoft/Recommenders • Contents • Utilities: modular functions for model creation, data manipulation, evaluation, etc. • Algorithms: SVD, SAR, ALS, NCF, Wide&Deep, xDeepFM, DKN, etc. • Notebooks: HOW-TO examples for end to end recommender building. • Highlights • 3700+ stars on GitHub • Featured in YC Hacker News, O’Reily Data Newsletter, GitHub weekly trending list, etc.
    [Show full text]
  • Alibaba Amazon
    Online Virtual Sponsor Expo Sunday December 6th Amazon | Science Alibaba Neural Magic Facebook Scale AI IBM Sony Apple Quantum Black Benevolent AI Zalando Kuaishou Cruise Ant Group | Alipay Microsoft Wild Me Deep Genomics Netflix Research Google Research CausaLens Hudson River Trading 1 TALKS & PANELS Scikit-learn and Fairness, Tools and Challenges 5 AM PST (1 PM UTC) Speaker: Adrin Jalali Fairness, accountability, and transparency in machine learning have become a major part of the ML discourse. Since these issues have attracted attention from the public, and certain legislations are being put in place regulating the usage of machine learning in certain domains, the industry has been catching up with the topic and a few groups have been developing toolboxes to allow practitioners incorporate fairness constraints into their pipelines and make their models more transparent and accountable. AIF360 and fairlearn are just two examples available in Python. On the machine learning side, scikit-learn has been one of the most commonly used libraries which has been extended by third party libraries such as XGBoost and imbalanced-learn. However, when it comes to incorporating fairness constraints in a usual scikit- learn pipeline, there are challenges and limitations related to the API, which has made developing a scikit-learn compatible fairness focused package challenging and hampering the adoption of these tools in the industry. In this talk, we start with a common classification pipeline, then we assess fairness/bias of the data/outputs using disparate impact ratio as an example metric, and finally mitigate the unfair outputs and search for hyperparameters which give the best accuracy while satisfying fairness constraints.
    [Show full text]
  • 2020 Sut20 CASC-J10.Pdf
    CASC-J10 CASC-J10 CASC-J10 CASC-J10 Proceedings of the 10th IJCAR ATP System Competition (CASC-J10) Geo↵Sutcli↵e University of Miami, USA Abstract The CADE ATP System Competition (CASC) evaluates the performance of sound, fully automatic, classical logic, ATP systems. The evaluation is in terms of the number of problems solved, the number of acceptable proofs and models produced, and the average runtime for problems solved, in the context of a bounded number of eligible problems chosen from the TPTP problem library and other useful sources of test problems, and specified time limits on solution attempts. The 10th IJCAR ATP System Competition (CASC-J10) was held on 2th July 2020. The design of the competition and its rules, and information regarding the competing systems, are provided in this report. 1 Introduction The CADE and IJCAR conferences are the major forum for the presentation of new research in all aspects of automated deduction. In order to stimulate ATP research and system de- velopment, and to expose ATP systems within and beyond the ATP community, the CADE ATP System Competition (CASC) is held at each CADE and IJCAR conference. CASC-J10 was held on 2nd July 2020, as part of the 10th International Joint Conference on Automated Reasoning (IJCAR 2020)1, Online, Earth. It was the twenty-fifth competition in the CASC series [139, 145, 142, 95, 97, 138, 136, 137, 102, 104, 106, 108, 111, 113, 115, 117, 119, 121, 123, 144, 125, 127, 130, 132]. CASC evaluates the performance of sound, fully automatic, classical logic, ATP systems.
    [Show full text]
  • Reverse Image Search for Scientific Data Within and Beyond the Visible
    Journal, Vol. XXI, No. 1, 1-5, 2013 Additional note Reverse Image Search for Scientific Data within and beyond the Visible Spectrum Flavio H. D. Araujo1,2,3,4,a, Romuere R. V. Silva1,2,3,4,a, Fatima N. S. Medeiros3,Dilworth D. Parkinson2, Alexander Hexemer2, Claudia M. Carneiro5, Daniela M. Ushizima1,2,* Abstract The explosion in the rate, quality and diversity of image acquisition instruments has propelled the development of expert systems to organize and query image collections more efficiently. Recommendation systems that handle scientific images are rare, particularly if records lack metadata. This paper introduces new strategies to enable fast searches and image ranking from large pictorial datasets with or without labels. The main contribution is the development of pyCBIR, a deep neural network software to search scientific images by content. This tool exploits convolutional layers with locality sensitivity hashing for querying images across domains through a user-friendly interface. Our results report image searches over databases ranging from thousands to millions of samples. We test pyCBIR search capabilities using three convNets against four scientific datasets, including samples from cell microscopy, microtomography, atomic diffraction patterns, and materials photographs to demonstrate 95% accurate recommendations in most cases. Furthermore, all scientific data collections are released. Keywords Reverse image search — Content-based image retrieval — Scientific image recommendation — Convolutional neural network 1University of California, Berkeley, CA, USA 2Lawrence Berkeley National Laboratory, Berkeley, CA, USA 3Federal University of Ceara,´ Fortaleza, CE, Brazil 4Federal University of Piau´ı, Picos, PI, Brazil 5Federal University of Ouro Preto, Ouro Preto, MG, Brazil *Corresponding author: [email protected] aContributed equally to this work.
    [Show full text]
  • Characterization and Prediction of Deep Learning Workloads in Large-Scale GPU Datacenters
    Characterization and Prediction of Deep Learning Workloads in Large-Scale GPU Datacenters Qinghao Hu1,2 Peng Sun3 Shengen Yan3 Yonggang Wen1 Tianwei Zhang1∗ 1School of Computer Science and Engineering, Nanyang Technological University 2S-Lab, Nanyang Technological University 3SenseTime {qinghao.hu, ygwen, tianwei.zhang}@ntu.edu.sg {sunpeng1, yanshengen}@sensetime.com ABSTRACT 1 INTRODUCTION Modern GPU datacenters are critical for delivering Deep Learn- Over the years, we have witnessed the remarkable impact of Deep ing (DL) models and services in both the research community and Learning (DL) technology and applications on every aspect of our industry. When operating a datacenter, optimization of resource daily life, e.g., face recognition [65], language translation [64], adver- scheduling and management can bring significant financial bene- tisement recommendation [21], etc. The outstanding performance fits. Achieving this goal requires a deep understanding of thejob of DL models comes from the complex neural network structures features and user behaviors. We present a comprehensive study and may contain trillions of parameters [25]. Training a production about the characteristics of DL jobs and resource management. model may require large amounts of GPU resources to support First, we perform a large-scale analysis of real-world job traces thousands of petaflops operations [15]. Hence, it is a common prac- from SenseTime. We uncover some interesting conclusions from tice for research institutes, AI companies and cloud providers to the perspectives of clusters, jobs and users, which can facilitate the build large-scale GPU clusters to facilitate DL model development. cluster system designs. Second, we introduce a general-purpose These clusters are managed in a multi-tenancy fashion, offering framework, which manages resources based on historical data.
    [Show full text]
  • A Universal Neural Network Solution for Tabular Data
    Under review as a conference paper at ICLR 2019 TABNN: A UNIVERSAL NEURAL NETWORK SOLUTION FOR TABULAR DATA Anonymous authors Paper under double-blind review ABSTRACT Neural Networks (NN) have achieved state-of-the-art performance in many tasks within image, speech, and text domains. Such great success is mainly due to special structure design to fit the particular data patterns, such as CNN captur- ing spatial locality and RNN modeling sequential dependency. Essentially, these specific NNs achieve good performance by leveraging the prior knowledge over corresponding domain data. Nevertheless, there are many applications with all kinds of tabular data in other domains. Since there are no shared patterns among these diverse tabular data, it is hard to design specific structures to fit them all. Without careful architecture design based on domain knowledge, it is quite chal- lenging for NN to reach satisfactory performance in these tabular data domains. To fill the gap of NN in tabular data learning, we propose a universal neural net- work solution, called TabNN, to derive effective NN architectures for tabular data in all kinds of tasks automatically. Specifically, the design of TabNN follows two principles: to explicitly leverage expressive feature combinations and to reduce model complexity. Since GBDT has empirically proven its strength in modeling tabular data, we use GBDT to power the implementation of TabNN. Comprehen- sive experimental analysis on a variety of tabular datasets demonstrate that TabNN can achieve much better performance than many baseline solutions. 1 INTRODUCTION Recent years have witnessed the extraordinary success of Neural Networks (NN), especially Deep Neural Networks, in achieving state-of-the-art performances in many domains, such as image clas- sification (He et al., 2016), speech recognition (Graves et al., 2013), and text mining (Goodfellow et al., 2016).
    [Show full text]