Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Building Models from Structured data with AutoML Tables

Building Models from Structured data with AutoML Tables

This is an introduction talk for AutoML Tables given at GDG Johannesburg Virtual Meetup

Peter-John Welcome

May 06, 2020
Tweet

More Decks by Peter-John Welcome

Other Decks in Technology

Transcript

  1. Peter-John Welcome GDE Firebase @pjapplez Building models from structured data

    with AutoML Tables An Introduction to AutoML Tables
  2. Half screen photo slide if text is necessary Who I

    am Mobile Engineering Lead at DVT Google Developer Expert for Firebase
  3. • Quantitative Data • Consistent field types • Easily searchable

    • Similar to relational databases or spreadsheets • Eg. Dates, Address, Coordinates, Money Structured Data
  4. • Qualitative Data • Inconsistent • More difficult to analysis

    and process • NoSQL Databases • Eg. Text, Audio, Images, Video Unstructured Data
  5. AutoML Tables Define your problem Retails Customer History: I am

    a retail company with customers and would like to predict how much money a customer will spend and what kind of products they are interested in. Conference Abstracts: I have a bunch of conference abstracts and I would like to predict if my abstract is likely to be accepted or reject.
  6. • Allows us to build and deploy ML models at

    speed and scale • Supports 3 types of Models: Binary classification, Multi-class classification, Regression Models • Has the ability to generate a REST API to access the model • Gives you the ability to Extract the model to test locally AutoML Tables
  7. • Training Data is split into 3 Categories: Training, Validation,

    Testing • By default its a 80(Training), 10(Validation),10 spilt(Testing) - This can be modified • Must have a minimum of 1000 rows • There is the ability to choose your optimization objectives AutoML Tables (Train)
  8. • Precision and Recall graphs • Confusion Matrix • Interactive

    Threshold score for precision and recall accuracy • Feature Importance AutoML Tables (Eval.)
  9. • Online Prediction • Batch predictions through BigQuery • REST

    endpoint deployment for client side interaction • Extraction of model to docker container to deploy in any environment AutoML Tables (Predict)
  10. • Easy to setup and get a model trained •

    Easy to deploy and lets Google Cloud scale the model for you • Model Size was really large • No way to extract the Model to TF model or TFLite model • Can get very expensive to Train large sets of data. • Can get expensive to deploy model. AutoML Tables (Observations)