AI Platform Unified Google Cloud
Ana López-Mancisidor ([email protected]) Teresa Muñoz-Cruzado ([email protected]) Vanessa Peinó ([email protected])
1
¿Por qué Google en Agenda 1 IA?
2 Enfoque IA en GCP
3 CCAI & Documen AI
4 ML & APIs
5 AI Platform Unified Proprietary + Confidential
¿Por qué Google en IA? Nuestras fortalezas en IA / ML
La organización de Contribuciones a la Calidad e Mejores prácticas investigación de IA comunidad de innovación en la de ingeniería para más grande del desarrolladores y producción empresas y MLOps mundo de código abierto (consumidor)
Cloud AI Platform Ejemplos de ML de Google en producción
Play Search App Recommendations Search Ranking Android Game Developer Experience Speech Recognition Keyboard and Speech Input
Drive Chrome Gmail Intelligence in Apps Search by Image Smart Reply
Spam Classification
YouTube Video Recommendations Maps Photos Better Thumbnails Street View Image Photos Search Parsing Local Search
Data Center Power Usage Ads Translate Reduced cooling Richer Text Ads Text, Graphic and energy 40% Automated Bidding Speech Translations
Self Driving Car
More than 1.5MM miles driven
Cloud AI Platform Nuestras fortalezas en IA / ML
Accesible Portabilidad Enterprise Seguridad
via API (TensorFlow, Ready ( los datos de los clientes son de los clientes, y Kubeflow) Google no usará dicha información para mejorar nuestros modelos ya entrenados o los modelos de otros clientes)
Cloud AI Platform Camino hacia la madurez en la IA
Táctica Estratégica Transformacional
Primeros casos de uso Varios sistemas desplegados y Aprendizaje continuo de ML y mantenidos en producción de ML MLOps Beneficios gracias a la mejora en la visibilidad de la información Modelos personalizados que Cultura centrada en el aprendizaje y permiten una ventaja competitiva experimentación continua
Cloud AI Platform Proprietary + Confidential
Enfoque IA en GCP Inteligencia Artificial en Google Cloud Invoca nuestras APIs: nuestros datos + nuestros modelos Modelos Construye tus propios modelos: Personalizados: tus datos + tu modelo tus datos + nuestro modelo Cloud Cloud Cloud Vision API Speech API Jobs API
Auto ML
Cloud Cloud Natural Cloud Video Translation API Language API Intelligence API
Cloud ML Engine Kubeflow Dialogflow
Cloud Speech Data Loss Synthesis API Prevention API Cloud AI portfolio Contact Center AI Document AI
ML APIs: pre-trained models with a single REST API Desarrolladores request de aplicaciones AutoML: train and serve your own models, no model code required
BQML: train and serve models with BQ data using a single SQL query
AI Platform: Notebooks, Hub, Pipelines, Training & Prediction
Kubeflow: deploy ML pipelines for pre-processing Modelos data, training, and serving models on Kubernetes personalizados Deep Learning VM images: spin up VMs with popular ML frameworks pre-installed
ML frameworks: TensorFlow, XGBoost, Sklearn, PyTorch, and more
Cloud AI Platform Proprietary + Confidential
CCAI What is Contact Center AI?
Automates basic chat and Makes human agents Unlocks insights about voice interactions more effective call drivers Dialogflow Agent Assist CCAI Insights Insights Understand metrics (omnichannel, across chat/voice/IVR)
Voice Dialogflow + telephony Real time prompts Virtual Human Agent assist agent Chat agent Dialogflow bot Dialogflow
Advanced conversational AI from google (TTS, STT, NLU) Understand ● Speech-to-text
Enabling a Talk conversation that ● Text-to-speech is close to human Interact ● Dialogflow ● Document assist ● Sentiment analysis Understand
Using the most advanced deep-learning neural network algorithms, Google Cloud’s Speech-to-text performs speech recognition withunparalleled accuracy.
10x 120+ ~5000 Larger vocabulary than the languages covered with more distinct phrase hints can be entire Oxford English Dictionary coming provided at real-time to customize the recognizer
~4 Formatting Real-time languages can be Real-life, context-specific Returns text as it’s recognized auto-detected formatting (e.g., $ or phone # and can analyze short & format) automatically added long-form audio Talk
With industry-leading voice quality, Google Cloud’s Text-to-speech enables a natural, conversational manner across channels.
180+ 92 21 Voices you can choose from Wavenet voices Languages covered with more coming
Wavenet REST/GRPC Pronunciation Exclusive multilingual access to Seamless integration with an Allows pauses, numbers, date & DeepMind’s WaveNet REST or GRPC capable time formatting, and other technology, which offer the application nuances most human-like voices Standard WaveNet
Try Text-to-Speech Try seamless, immediate, with any text! speech-text-speech translation! Interact Reaching +1Mn developers, Dialogflow is the emerging standard for building natural and rich conversational experiences across multiple channels.
1M+ 32 17 developers on the platform Languages and variances with Single Click integrations with more coming popular chat and messaging platforms
2000 REST/GRPC Flexible Intents. Speak to us if you need HIPAA and PCI Compliant and Platform agnostic built on GCP, more. similar SLO than other Cloud with Cloud Support and SLA products available. 7 SDKS to to integrate in your applications (python, node, java, Go, etc)
Proprietary + Confidential
Document AI Document AI extracts & classifies information from unstructured documents
01 02 03
Read it Understand it Make it useful Life of a document
CATEGORIZE CONTENT DETECT DIAGRAMS INGEST & FILTER Categorize patent’s content Identify diagram and Document is read in from using NLP model. corresponding x and y Cloud Storage. coordinates. Unstructured document, multiple formats and languages
OCR EXTRACT ENTITIES STORAGE Extract out raw text into Identify named entities in Write out and store json format for the raw text. results from the pipeline downstream NLP process. into BigQuery Demo: https://cloud.google.com/solutio ns/document-ai What is Document AI?
DocAI turns Document image unstructured content
{Class: /us/gov/ID/Driverslicense into structured data. State: Ohio Name: Jane Doe Issued: 7/2/2018 Expires: 1/21/2021 }
Structured data Invoice Parser I18n: English, French, Dutch, German, Spanish; 150+ more languages by March 2021 Proprietary + Confidential
ML APIs Machine Learning APIs: Ready to Go
Cloud Cloud Vision API Speech API
Cloud Cloud Natural Cloud Video Translation API Language API Intelligence API Vision API
Object recognition Logo and facial analysis Text extractions Detect inappropriate content Natural Language API
Classify content Detect sentiment Extract entities Analyze syntax Demo time! https://cloud.google.com/vision/docs/drag-and-drop Natural Language API: Sentiment Analysis
Inspects the given text and identifies the prevailing emotional opinion within the text, especially to determine a writer's attitude as positive, negative, or neutral. ● Score: positive, negative, neutral
● Magnitude: indicates the overall strength of emotion (both positive and negative) within the given text Demo: Analyze the sentiment of user opinions with Google Cloud AI
Fills Feedback Form
Analyze Sentiment Opinion User Portal Feedback Form
Natural Language API - Updates Survey Ejecuta cálculos Sentiment Analysis
Cloud Runs queries for Functions aggregated data
Portal Aggregated Data Alerts (Pub/Sub) Administración Survey
BigQuery
Report Updates data for visualization analysis ML Frameworks: AutoML: Machine Learning APIs: Total Control Bring Your Own Data Ready to Go (We Do the Rest)
Cloud Cloud Vision API Speech API TensorFlow AI Platform
Cloud Cloud Natural Cloud Video Spark ML Cloud Dataproc Translation API Language API Intelligence API light
light
chair chair sofa
table light mod.A
light mod.B
chair mod.1 chair mod.2 sofa mod.Y
table mod.Z Data Tune ML model preprocessing ML model design Evaluate Deploy Update parameters Data Tune ML model preprocessing ML model design Evaluate Deploy Update parameters
Large computational resources Machine learning expertise Dataset Cloud AutoML Generate predictions with a REST API
Train Deploy Serve AutoML Vision
Upload and label images Train your model Evaluate
Cloud AutoML Handbag Shoe Hat
AutoML Natural Language
Upload and label text Train your model Evaluate
Sports
Lifestyle
Money
Tech Cloud AutoML Travel
Sports Lifestyle, Tech Money
AutoML Tables
Upload labeled Train your model Evaluate (Classification structured data / Regression)
ID Feature Feature ... ID Feature Feature ... Label Class A ...... Class B ... Class C ... Cloud AutoML Class D
Model is now trained and ready to make predictions This model can scale as needed to adapt to customer demands AutoML Edge
Deploy to edge devices
Ingest data Use Google’s Use AutoML Vision Export Data Labeling Edge to train your high-accuracy, Service if needed model and optimize low latency model it for desired latency AutoML Edge
Easily build and deploy high-accuracy, low-latency models to edge devices
Classify images on-device Achieve fast, Supports a variety of Ensure privacy by keeping and trigger actions in state-of-the-art edge devices data on the device, never real-time, even with limited prediction accuracy for a leaving your enterprise or unreliable connectivity range of applications premise Demo time! Comprehensive Suite for AI Developers
Sight Language Conversation Structured Data
Dialogflow Enterprise Cloud Vision Cloud Translation AutoML Tables Edition
Cloud Video Cloud Natural Recommendation AI Cloud Text-to-Speech Intelligence Language (now GA!) Cloud AutoML Vision AutoML Translation Speech-to-Text Cloud Inference API (alpha) AutoML Video AutoML Natural Intelligence Language Summary
Cloud AutoML offershigher model accuracy and faster time to create a production-ready model.
Users need limited ML expertise to build a custom
ML model.
Simple graphical user interface to upload their images, train and manage their models, and then use those models directly on Cloud. Customer asks...
Can we identify relevant objects from an image?
Vision API / AutoML Vision Customer asks...
Can we predict the price of a house based on different attributes?
AutoML Tables (regression) Customer asks...
Can we personalize recommendations for our theme park visitors based on demographics, weather, etc ?
Recommendations AI Customer asks...
Can we extract patient data (name, insurance, etc) from a form automatically?
Document AI / AutoML Vision Customer asks...
Can we monitor financial product selling phone calls to identify if our agents are elaborating on our key products?
Speech to Text / Text Classification Customer asks...
Voice ordering assistant for our fast food mobile app channel?
Dialogflow Today’s demos
Vision API using the API Explorer API Explorer: https://developers.google.com/apis-explorer Vision API method images.annotate: https://cloud.google.com/vision/docs/reference/rest/v1/images/annotate?apix=true Vision API feature types: https://cloud.google.com/vision/docs/reference/rest/v1/Feature#type
Natural Language API Python demo Natural Language API client libraries: https://cloud.google.com/natural-language/docs/reference/libraries Natural Language API code samples: https://cloud.google.com/natural-language/docs/samples
AutoML Vision demos Cloud-hosted model quickstart:https://cloud.google.com/vision/automl/docs/quickstart Edge device model quickstart:https://cloud.google.com/vision/automl/docs/edge-quickstart Export edge model:https://cloud.google.com/vision/automl/docs/export-edge (TF Lite + android, Edge Container) Quick API demos
Document AI https://cloud.google.com/document-ai/docs/drag-and-drop
Translate https://cloud.google.com/translate
Vision AI https://cloud.google.com/vision
Natural Language https://cloud.google.com/natural-language
Speech to Text https://cloud.google.com/speech-to-text API & AutoML qwiklabs
Image Processing
Analyzing a Tennis Serve with the Video Intelligence API
Building a Document AI Pipeline with Google Cloud
Integrate with Machine Learning APIs
Classify Images of Clouds in the Cloud with AutoML Vision Proprietary + Confidential
BigQuery ML Proprietary + Confidential
Unique
Fully managed and serverless Google Cloud’s enterprise data for maximum agility and scale warehouse for analytics Unique
Real-time insights from streaming data Exabyte-scale storage and Petabyte-scale SQL queries Google Unique BigQuery Built-in ML for out-of-the-box predictive insights Encrypted, durable, and highly available Unique
High-speed, in-memory BI Engine for faster reporting and analysis Proprietary + Confidential
BigQuery ML
Execute ML initiatives without moving data from BigQuery
Iterate on models in SQL in BigQuery to increase development speed
Automate common ML tasks
In-built infra management, security & compliance Proprietary + Confidential
Supported models
Logistic regression k-means clustering Classification Other Models
DNN classifier (TensorFlow) Time series forecasting Boosted trees using XGBoost Recommendation: AutoML Tables Matrix factorization
Linear regression TensorFlow models for Regression Model batch and online DNN regressor (TensorFlow) Import/Export prediction Boosted trees using XGBoost
AutoML Tables Architecture
Collect and Transform Predict Visualize Activate
Data Scientists Data Studio Business Analysts Executives BigQuery GA360 Export BigQuery BigQuery Storage ML
GA360 GMP Proprietary + Confidential
SQL interface to Machine Learning
● Data scientists can: ○ Build 100 models in a week ○ Not spend time on data ETL and setup of multiple tools ○ Experiment in BigQuery, export models to CAIP for further tuning ● Data analysts can: ○ Build ML models with knowledge of basic ML concepts. ○ No need to use multiple languages, just SQL ○ Easily utilize domain and data knowledge for ML models ● All: model can be served immediately after it’s trained CREATE MODEL `demo.return_buyer_model` OPTIONS (model_type='automl_classifier', AutoML labels = ['will_buy_later']) AS SELECT ... tables through
BigQuery ML SELECT fullVisitorId, predicted_will_buy_later, predicted_will_buy_later_probs FROM ML.PREDICT(MODEL `demo.return_buyer_model`, (SELECT ... Proprietary + Confidential
Export to GCS
PROJECT=$(gcloud config get-value project) BUCKET=${PROJECT}-eu gsutil mb -l eu gs://${BUCKET} bq extract -m ch09eu.bicycle_model_linear \ gs://${BUCKET}/bqml_model_export/bicycle_model_linear
Deploy to Cloud AI Predictions
./deploy.sh gs://${BUCKET}/bqml_model_export/bicycle_model_linear \ Online Prediction europe-west1 london_bicycles bqml
Run Online Predictions
gcloud ai-platform predict --model london_bicycles \ --version bqml --json-instances input.json Demo time! Proprietary + Confidential
AI Platform AI Platform (unified) Proprietary + Confidential
● End-to-end, code-based development environment for AI inside GCP console built using expertise from Google’s internal AI development
● Individual AutoML products will be integrated as features into Cloud AI Platform to create a cohesive workbench for ML builders
● Enterprise customers will benefit from a seamless, cutting-edge ML development experience from Our AI platform enables customers to build AI applications for their own organization leveraging beginner to expert and for any ML use Google’s technology, scale and “way of doing things”. case Proprietary + Confidential
AI Platform Applications
Vision and Video Conversation Language Structured Data
Core
Notebooks Data Labeling Deep Learning Env Experiments ML Metadata
AutoML Training Explainable AI Feature Store Vizier (Optimization)
Prediction Model Monitoring ML Pipelines AI Accelerators Hybrid AI A unified object framework across AI Platform Proprietary + Confidential
● Unstructured or structured datasets housed within GCP with managed Dataset metadata including annotations ● Can be stored in GCS object storage, CSVs or BQ Tables
Training Pipeline ● Objected created to train a model using a dataset ● Can also use external datasets in compatible formats
● Model Machine learning model artifact with metadata built with a Training Pipeline ● Compatible models can also be uploaded directly to the platform
● Endpoint that can be called by users for online predictions and explanations Endpoint ● Contains one or more deployed models which may exist independently from other model artifacts AI Platform Notebooks Proprietary + Confidential Features Proprietary + Confidential
● Schedule execution of notebooks ● Integration with GCP: BigQuery and GCS data exploration, execution with CAIP training ● Ability to scale hardware up or down directly from JupyterLab. Features Proprietary + Confidential
● Secure development: private networking, VPC-SC, shared VPC and private IP controls. You can also encrypt your data on disk with CMEK. ● Controlled user access: single-user or service account ● Support for data science frameworks, including R, pandas, NumPy, SciPy, scikit-learn, and Matplotlib, and ML frameworks like TensorFlow, Keras, fast.ai, RAPIDS, XGBoost, and PyTorch. query=""" SELECT weight_pounds, is_male, mother_age, Download plurality, gestation_weeks BigQuery data to FROM publicdata.samples.natality our notebook WHERE year > 2000 LIMIT 10000 """ df = bigquery.Client().query(query).to_dataframe() df.head() Demo time! AI Platform prices - AutoML Proprietary + Confidential
For AutoML models, you pay for three main activities: ● Training the model ● Deploying the model to an endpoint ● Using the model to make predictions AI Platform prices - Training and Prediction Proprietary + Confidential
Price per node hour, depending on type machine, accelerator and disk type. Example for prediction: AI Platform prices - Data Labeling Proprietary + Confidential
Price per 1,000 units per human labeler (Tier-1, tier-2 are volumes) Proprietary + Confidential
Resources
AI Platform Product Page cloud.google.com/ai-platform
ML Engineer Certification cloud.google.com/certification/machine-learning-engineer
TensorFlow Extended tensorflow.org/tfx
KubeFlow Pipelines kubeflow.org/docs/pipelines/ Gracias
Confidential & Proprietary