Google Dev Fest 2018

Google Dev Fest 2018



  • Study Jams?
  • A little unorganized, rough transitions between events, etc.

Keynote: Actions on Google - Daniel Myers

  • Extends google assistant to do things via conversation
  • (I need to work on my presentation skills)
  • Major shift in computing every 10 years
    • Mainframe -> Desktop -> Internet -> Mobile -> AI/Assistants
  • Google assistant + actions + devices w/ assistant built in
  • Assistant uses speech to text nlp knowledge graph ml ranking user profile to detect which action to invoke
  • Regex just isn’t sufficient to parse the text (“It’s kind of cold outside, so I’d like something to warm me up like a hot soup)
  • Dialog Flow
    • Intent matching
      • Match and categorize utterances to an intent
    • Entity extraction
      • Identify key words and phrases spoken by user
  • Can you create a dialog flow with a file for source control? Or is it all GUI? Is there an export?
  • System Entities (Time, Speed, Weight, Age, Currency, etc.)
  • Custom entities (milk type, protein, etc.)
  • Specify follow-up questions if user doesn’t specify certain values
  • Fulfillment
    • Where you actually write your code
    • Big Json document for the conversations, including original text and what the system derived
  • Integrations with almost all major chat services
  • - easiest way to interact with actions on google
  • Guide the user (suggestion tips)
  • Display basic cards if a screen is available
  • Lists, carousels and browsing carousels for selection
  • Ask for information (like location)
  • Account linking
  • Transact with the user
    • Use your payment processor
    • GPay
  • Smart Home Device Integration
  • Sample Actions
    • Use it as a foundation and expand on it
  • Actions on Google Community Program

Mobile Machine Learning (Tensorflow) - Laurence Moroney

  • Automatic closed captioning
    • No punctuation at all
  • A lot of people afraid of AI progress
  • Replacing AI with Demonic Possession ad agency in NY
  • Gartner Hype Cycle
    • Tech Trigger -> Inflated expectations -> Trough of disillusionment -> Enlightenment -> Productivity
  • If we can “tunnel through the hype” to disillusionment, we can actually progress towards enlightenment
  • On-ramp to artificial intelligence is Machine Learning
  • Deep learning is Machine Learning using neural networks
  • Traditional Programming
    • Rules + Data -> Program -> answers
  • ML
    • Answers + Data -> Program -> Rules
  • Rules become infeasibly difficult to manage (walk, run, bike, golf)
  • Neural network - horribly named technology
    • functions that pass their output to other functions
  • Predictions
    • Probability that data is associated with a rule
  • Code
    • keras

Federated Machine Learning - Emily Glanz

  • FL, not FML
  • Hype Video - Making every phone smarter with federated machine learning
  • learn from other devices while keeping your data private
  • Privacy is the default, allows device to learn from other users’ usage patterns
  • What’s a device
    • Phone - Average adult is on phone for 10 hours a day in USA
    • IoT/Edge Devices
    • Robots - actively interact/learn from devices

Growing with Google

  • Scholarship program
  • Training, tools & events
  • Google Developers Training
  • Google Developer Scholarship Program

Martin Omander - Google Cloud Functions

  • Deploying infrastructure is a common procedure
  • GCP can pull from things like google spreadsheet just by sharing with the GCP account - kinda cool
  • select joke from jokes order by rand() limit 1
  • reads only one row
  • from microservice writing to pub-sub that inserts into data analytics
  • writing to pub-sub to allow different parts of the service to subscribe to that service
  • costs
    • Run time 200 ms/function
    • Memory 256 mb
    • Processor 400MHz
    • Cloud SQL - $8-55/month
    • Cloud Datastore (NoSQL): $0-6/month

KubeFlow - Amy Unruh

ML Stack on top of Kubernetes

  • ML code is a tiny part of the actual ML workflow
  • Setting up a production ML stack/pipeline is super hard
  • Kubeflow is designed to tackle a lot of this
  • Make it easy for everyone to develop, deploy and manage portable distributed ml on kubernetes
  • Who
    • Data Scientists, ml resarch, software engineer, etc.
  • What
    • Portable ML products on k8s
  • Why
    • Because building a platform is too big of a problem to tackle alone
  • Can deploy to any kubernetes conformant cluster (on any cloud provider)