What is AI?

AI > Machine Learning > Deep Learning

From Ad-how Data Analysis to Data-Driven Devisions

Ad-How:

• Manual, repetitive work
• Any decision is based on hunches how factors interact

=> Build forecasting ML Models

Options for ML Models

Notebooks

Customize hardware for your specific notebook needs.

BigQuery Magic

BigQuery result is saved in pandas dataframe ‘df’. Be aware that memory is limited in notebooks.

%%biquery df
select
  *
from
  ...
limit
  50

ML on Unstructured Data

Unstructured data is about 90% of a companies data.

ML can automate tasks.

Different approaches to AI. You need 100k+ Datapoints to train your own model.

Prebuild Models

Natural Language API

Enitity analysis

• Locations
• People
• Artwork
• Consumer Products
• Addresses
• Numbers (Prices, Phone numbers, Dates ..)

Sentiment Analysis

Entity Sentiment Analysis

Sentiment analysis on each eantity of a document.

Content Classification

• Sports
• Foods
• Literature

Vision API

Translate API

Language API

Sentiment Analysis

Speech API

Video Intelligence API

Dialog Flow

Natural Language understanding API.

Prebuilt Chatbots

AutoML

• Train
  • Dataset preparation
  • Readiness Analysis
  • Train
  • Test and Evaluate
• Deploy
• Server
  • Predict
  • Classify

Process

Prepared Dataset to train custom Model

Preliminary Analysis

One full run through all training data is called ‘epoch’.

Evaluation

Activate Model

Serve Model

Multiple Models

AutoML Products

AutoML Vision

AutoML takes more time to come up with a model, due to ensemble learning.

Attribute	AutoML Vision	Vision API
Objective	Enabling developers with no ML expertise to build ML models	Enable ML practitiones to use Googles ML
Primary use	Classification	Face detection, OCR …
Data	Images with Labels	Justt Images

Process

Improving Data

AutoML Natural Language

Process

Improving Data

AutoML Tables

Price Suggestions

Custom Models

• BigQuery ML
• TensorFlow, Keras

Model Quality: Custom Model > AutoML > BigQueryML

Make or Buy Models?

BigQuery ML, AutoML or Custom Model?