Upload
india-quotient
View
122
Download
0
Embed Size (px)
Citation preview
Cloud Machine Learning Google Cloud Platform
[email protected]( Business Development )
[email protected] (Customer Engineer)
Data is exploding.And smart companies are taking advantage.
Unstructured data accounts for 90% of enterprise data*
Cloud Machine Learning help you make sense of it
*Source: IDC
Confidential & ProprietaryGoogle Cloud Platform 4
What is Machine Learning?
Data Algorithm Insight
Confidential & ProprietaryGoogle Cloud Platform 5
Machine Learning @ Google
Beach
Woman
Pool
Coast
Water
Confidential & ProprietaryGoogle Cloud Platform 7
Google Translate
Confidential & ProprietaryGoogle Cloud Platform 8
Confidential & ProprietaryGoogle Cloud Platform 9
Enterprise Predictive Analytics Challenges
Data access to a variety of data sources.
Develop and build analytic models.
Data preparation, exploration and visualization.
Deploy models and integrate them into business processes
and applications.
High performance and scalability for both development
and deployment.
Perform platform, project and model management.
Confidential & ProprietaryGoogle Cloud Platform 10
Data Warehouse is the Foundation of Something Bigger
Data Warehouses/Lakes
Machine Intelligence Predictive +
Prescriptive Analytics
=Advanced Analytics
Cloud
On Premises
MachineLearning
APIs
Train Your Own
Models
Confidential & ProprietaryGoogle Cloud Platform 11
Machine Learning Use Cases
• Predictive maintenance or condition monitoring• Warranty reserve estimation• Propensity to buy• Demand forecasting• Process optimization
Manufacturing
• Predictive inventory planning• Recommendation engines• Upsell and cross-channel marketing• Market segmentation and targeting• Customer ROI and lifetime value
Retail
• Alerts and diagnostics from real-time patient data• Disease identification and risk satisfaction• Patient triage optimization• Proactive health management• Healthcare provider sentiment analysis
Healthcare and Life Sciences
• Aircraft scheduling• Dynamic pricing• Social media – consumer feedback and interaction analysis• Customer complaint resolution• Traffic patterns and congestion management
Travel and Hospitality
• Risk analytics and regulation• Customer Segmentation• Cross-selling and up-selling• Sales and marketing campaign management• Credit worthiness evaluation
Financial Services
• Power usage analytics• Seismic data processing• Carbon emissions and trading• Customer-specific pricing• Smart grid management• Energy demand and supply optimization
Energy, Feedstock and Utilities
Confidential & ProprietaryGoogle Cloud Platform 12
Why So Little Machine Learning Apps Out There?
• Building and scaling machine learning infrastructure is hard
• Operating production ML system is time consuming and expensive
Confidential & ProprietaryGoogle Cloud Platform 13
Building Smart Applications Today
Technology Operationalization Tooling
Difficult to scale
Many choices for different use cases
Using latest technology (e.g. DNN) is hard
Complex data pipelines
Managing ML infra takes away time from actually doing ML
Many models to manage
Complex dev pipeline with many combinations of tools/libraries
Not fully interactive developer experience - collaboration/sharing is hard
Confidential & ProprietaryGoogle Cloud Platform 14
Introducing Cloud Machine Learning
● Fully managed service
● Train using a custom TensorFlow graph for any ML use cases
● Training at scale to shorten dev cycle
● Automatically maximize predictive accuracy with HyperTune
● Batch and online predictions, at scale
● Integrated Datalab experience
Confidential & ProprietaryGoogle Cloud Platform 15
Cloud Datalab
● Interactively explore data
● Define features with rich visualization support
● Launch training and evaluation
● ML lifecycle support
● Combine code, results, visualizations & documentation in notebook format
● Share results with your team
● Pick from a rich set of tutorials & samples to learn and get started with your project
Confidential & ProprietaryGoogle Cloud Platform 16
Powerful Machine Learning Algorithm
● Convolutional Neural Network for image classification
● Recursive Neural network for text sentiment analysis
● Linear regression at scale to predict consumer action (purchase prediction, churn analysis)
● And unlimited variety of algorithms you can build using TensorFlow
Confidential & ProprietaryGoogle Cloud Platform 17
Automatically tune your model with HyperTune
● Automatic hyperparameter tuning service
● Build better performing models faster and save many hours of manual tuning
● Google-developed search algorithm efficiently finds better hyperparameters for your model/dataset
HyperParam #1
Obje
ctive
Want to find this
Not these
HyperParam #2
Confidential & ProprietaryGoogle Cloud Platform 18
Integrated with GCP Products
● Access data that is stored in GCS or BigQuery
● Save trained models to GCS
● Preprocess largest datasets (TB) using Dataflow
● Orchestrate ML workflow as a Dataflow pipeline
● Analyze data and interactively develop ML models in Datalab
Confidential & ProprietaryGoogle Cloud Platform 19
Fully Managed Machine Learning Services
● Scalable and distributed training infrastructure for your largest data sets
● Scalable prediction infrastructure that can serve very large traffic
● Managed no-ops infrastructure handles provisioning, scaling, and monitoring so that you can focus on building your models instead of handling clusters
Confidential & ProprietaryGoogle Cloud Platform 20
Pay As You Go and Inexpensive
Tier Price
Regular $0.1 / 1K +$0.40/Node Hour
Large volume $0.05/1K +$0.40/Node Hour after 100M/month
Training PredictionUS Europe / Asia
1 ML training unit $0.49 $0.54
Tier Training unit per hour
Characteristics
BASIC 1 A single worker instance. This tier is suitable for learning how to use Cloud ML, and for experimenting with new models using small datasets.
STANDARD_1
10 Mid-size cluster with many workers and a few parameter servers for medium scale distributed training
PREMIUM_1
75 Larger cluster with a large number of workers with many parameter servers. Suitable for large scale job with complex and larger models
CUSTOM Custom Fine tune the number of workers, parameter servers and machine types
Do You Have The Right Visibility?
** Ventura Research Report** Ventura Research Report
34% 51%
71%
Of retail companies are satisfied with the processes they use to
create analytics.
Of retailers are still using spreadsheets as their primary
data analysis tools
Find challenge in data sharing
45%
Are not effectively using data to personalize
marketing communications 42%
Are not able to link data together at the individual
customer level
Largest ObstacleRetail Analytic Trends
Challenges
Difficulty Understanding Customers
What drives the customers buying habits?
What products do customers prefer to buy and what related products?
What causes customers to not buy?
Customizing The Experience
How can I ensure each customer sees the products they’re interested in as quickly as possible?
How can my eCommerce app react in real-time to customer actions?
Data Aggregation & Processing
Need for a large, scalable storage solution to aggregate, store, and serve applications
Compute capacity required to churn and derive insights constantly increasing
Analytics & Machine Learning can be resource hogs
Key Takeaways Data is a core business assetAnalytics drive competitive advantageData at scale drives exponential complexity
Traditional BI does not scale to big dataMost organizations cannot capture all dataInformation growing faster than it can be leveraged
Retail Drivers - How Analytics Can Help?
DemandingCustomers
AggressiveCompetition
CostOptimization
ImproveExperience
UnderstandCustomers
FasterConversions
IncreaseSales
Customer Profiling Segmentation
Recommendations Cart Analysis
Market Hot Spotting
AssetPerformance
Social Media Analysis
Customer Personalization
Data AggregationMultiple Platforms
Location Planning
Catchment Analysis
Inventory Management
Logistics Management
Sales Forecasting
Impact Analysis
Risk Modeling
Confidential & ProprietaryGoogle Cloud Platform 24
Transform Data into Actions
Exploration & CollaborationDatabases Storage
Data Preparation &
Processing Analytics
Advanced Analytics & Intelligence
Mobile apps
Sensors and devices
Web apps
Relational
Key-value
Document
SQL
Wide column
ObjectStream processing
Batch processing
Data preparation
Federated query
Data catalog
Data exploration
Data visualization
Developers
Data scientists
Business analysts
Development environment for Machine
Learning
Pre-Trained Machine Learning models
Data Ingestion
Messaging
Logs
Confidential & ProprietaryGoogle Cloud Platform 25
Transform Data into Actions
Data Preparation &
Processing
Cloud Dataflow
Cloud Dataproc
Exploration & Collaboration
Google BigQuery
Cloud Datalab
Google Analytics 360
Cloud Dataproc
Mobile apps
Sensors and devices
Web apps
Developers
Data scientists
Business analysts
Data Ingestion
Cloud Pub/Sub
App Engine
Databases/Storage
Cloud SQL
Cloud Bigtable
Cloud Datastore
Cloud Storage
Analytics
Google BigQuery
Google Analytics 360
Cloud Dataproc
Google Drive
Advanced Analytics & Intelligence
Cloud Machine Learning
Translate API
Vision API
Speech API
Confidential & ProprietaryGoogle Cloud Platform 26
Use Your Own Data to Train Models
BETA
BETA
GAGA
Cloud Datalab
Cloud Machine Learning
Cloud Storage Google BigQuery Develop/Model/Test
Confidential & ProprietaryGoogle Cloud Platform 27
HTTP request
Use your own data to train models
Pre-ProcessingData Storage
Training flow
Prediction flow
Localtraining
Download
Mobileprediction
Batch
Online
Training
Prediction
Tooling
Datalab
Datalab
Tooling
UploadHosted Model
Confidential & ProprietaryGoogle Cloud Platform 28
Automatically categorize, and automatically extract value
Evaluate the model by applying it against
additional manually categorized data, correct
and tune
Capture thousands of examples of correct evaluations for that
categorization, and use them to train an ML model
Identify categorizations that provide value, categories you’re
already evaluating for by hand today
1 2 3 4
Machine Intelligence is Already Making a Huge Difference and There are Many, Many More Opportunities
Confidential & ProprietaryGoogle Cloud Platform 29
Machine Learning @ GoogleLevel 200
Confidential & ProprietaryGoogle Cloud Platform 30
The point of ML is to make predictions
Input Feature Predicted Value
Model
Confidential & ProprietaryGoogle Cloud Platform 31
Tensorflow helps you “train” models
Input Feature Predicted Value
Model
True ValueUpdate model based on Cost
Cost
Confidential & ProprietaryGoogle Cloud Platform 32
Democratizing machine learning
App DeveloperData Scientist
CloudML
Build custom modelsUse/extend OSS SDK
Scale, No-ops Infrastructure
ML APIs
Vision API
Speech API
Use pre-built models
Translate API
ML researcher
Language API
Confidential & ProprietaryGoogle Cloud Platform 33
Beyond Tensorflow
Size of datasetSize of NN
Scale of Compute Problem
Accuracy
CloudML ( )
Deep networks
TensorFlow Processing Units (TPUs)
Distributed
No-ops
https://cloudplatform.googleblog.com/2016/05/Google-supercharges-machine-learning-tasks-with-custom-chip.html
ML APIs
Vision API
Speech API
Translate API
Language API
Confidential & ProprietaryGoogle Cloud Platform 34
ML APIs are simply REST calls and can be made from any language or framework
sservice = build('speech', 'v1beta1', developerKey=APIKEY)response = sservice.speech().syncrecognize( body={ 'config': { 'encoding': 'LINEAR16', 'sampleRate': 16000 }, 'audio': { 'uri': 'gs://cloud-training-demos/vision/audio.raw' } }).execute()print response
Data on Cloud Storage
cloud.google.com