7. Train your models at scale
Cloud Machine
Learning Engine
● Write your model using the Dataset and Estimator APIs
● Stage it in Cloud Storage
● Store training and evaluation datasets in Cloud Storage
● Configure your cluster (CPU, memory, GPU/TPU) — can
also run in local mode
● Write your checkpoints to Cloud Storage
● Tune the hyperparameters of your model1
● Monitor from the Google Cloud Console, Stackdriver
Logging, TensorBoard and the command-line
1
http://bit.ly/mle-hypertune
8. Host your trained model on the cloud
● Online prediction with serverless, fully-managed hosting
● Batch prediction at scale
● Manage deployed models and versions
● Restrict and audit access to your deployed models
● Monitor your predictions in Stackdriver Logging
Cloud Machine
Learning Engine
23. AI models optimized
for Edge
Training
& inference
AI
TensorFlow Lite
& Android NN API
Cloud IoT Edge Cloud AI & servicesSoftware
Edge Cloud
CPUs, GPUs
Hardware
CPUs, GPUs
Cloud TPU
AI models
Edge TPU
27. BigQuery ML
● Petabyte-scale, fast linear and binary logistic regression
● SQL-like language:
○ CREATE MODEL
○ ML.EVALUATE, ML.ROC_CURVE
○ ML.PREDICT
● Automatic learning rate adjustment
● L1
and/or L2
regularization
● Batch prediction within BigQuery
● Exportable models1
:
○ ML.WEIGHTS
○ ML.FEATURE_INFO
BigQuery
NEW
1
http://bit.ly/bqml-online
28. Training a model
CREATE MODEL `bqml_tutorial.natality_model`
OPTIONS (model_type='linear_reg',
input_label_cols=['weight_pounds']) AS
SELECT weight_pounds,
is_male,
gestation_weeks,
mother_age,
CAST(mother_race AS STRING) AS mother_race
FROM `bigquery-public-data.samples.natality`
WHERE weight_pounds IS NOT NULL
BigQuery
29. SELECT * FROM
ML.EVALUATE(
MODEL `bqml_tutorial.natality_model`, (
SELECT weight_pounds,
is_male,
gestation_weeks,
mother_age,
CAST(mother_race AS STRING) AS mother_race
FROM `bigquery-public-data.samples.natality`
WHERE weight_pounds IS NOT NULL)
)
BigQuery
Returns: mean_absolute_error, mean_squared_error, mean_squared_log_error,
median_absolute_error, r2_score, explained_variance
Evaluating a model
30. Getting predictions from a model
SELECT predicted_weight_pounds FROM
ML.PREDICT(
MODEL `bqml_tutorial.natality_model`, (
SELECT is_male,
gestation_weeks,
mother_age,
CAST(mother_race AS STRING) AS mother_race
FROM `bigquery-public-data.samples.natality`
WHERE state = "WY")
)
BigQuery
44. ● Image properties (dominant colors, crop hints)
● Landmark recognition
● Handwriting recognition
● Object detection
● General availability
● Additional file types: PDF and TIFF
Cloud Vision API
Other features
NEW
NEW
NEW
NEW
45. Cloud
Speech-to-Text
Cloud Video
Intelligence API
Cloud Natural
Language API
Cloud
Translation API
● Language identification
● Word-level confidence scores
● Multi-participant recognition
Cloud
Text-to-Speech
● DeepMind WaveNet voices
● Output channel optimization
Other pretrained models
NEW
NEW
NEW
NEW
NEW
47. ● Based on transfer learning and neural architecture search
● Prediction available from REST APIs
● For (beta): domain-specific translated pairs
● For (beta): predict domain-specific
categories (single or multi-label classification)
● For (beta): detect custom objects or predict
domain-specific labels
Cloud Auto ML
Add your domain-specific knowledge
50. ● Phone Gateway (beta): assign a phone number to a virtual
agent, with speech recognition, speech synthesis and
natural language understanding
● Knowledge Connectors (beta): understand unstructured
documents like FAQs or knowledge base articles and
complement your pre-built intents
● Automatic Spelling Correction (beta)
● Sentiment Analysis (beta)
● Text-to-Speech (beta)
● Enables
DialogFlow
Enterprise Edition
Interact via natural language
NEW