Module 1&2

  • Model: math function
  • Supervised Learning: adjust model to match output and label

input, label –> model –> output

  • ML have two stages: training and inference
    • Both are important for ML be useful.
    • Inference takes inputs that model has not seen.
  • replace heuristics with model
  • starts with heuristics but use ML when have data

Use case questions:

  • If the use case was an ML problem….
    1. What is being predicted?
    2. What data is needed?
  • Now imagine the ML problem is a question of software:
    1. What is the API for the problem during prediction?
      • Inputs & Outputs
    2. Who will use this service? How are they doing it today?
  • Cast it in the framework of a data problem. What are some key actions to collect, analyze, predict, and react to the data/predictions
    1. What data are we collecting?
    2. What data are we analyzing?
    3. What data are we predicting?
    4. What data are we reacting to?

ML is scaling beyond hand-written rules. Simple model + more data > Complex model + less data

  • Don’t jump into ML! Start with manual data analysis to know your data!
Training-serving skew
  • Batch data in training v.s. streaming data in prediction
  • The pipeline should handle both to get the same input for the model.

Fail fast and iterate!

ML business (Module 3)

  • Defining KPI’s
  • Collecting data
  • Building infra
  • Optimization

  • Integration

  • Effort Allocation
    • Top1: Collecting data: to gather more good-quality data to serve this problem.
    • Top2: Building infra: make sure to serve smoothly to users.
    • Top3: Integration
  • What is the application for?
    • Is your ML is improving the real world?

ML value comes along the way.

The generic feed-back loop
st=>start: inputs 
e=>end: done
process=>operation: Process
output=>operation: Output
ig=>operation: Insight generation (Big Data)
tune=>operation: Tuning (ML)
cond=>condition: Good enough?

st->process->output->ig->tune->cond
cond(yes)->e
cond(no)->process
  • Process:
    • Individual: prototyping, trying new ideas.
    • Delegation: Gently increase people involved, formalize the process.
    • Digitization: Automate repetitive parts. Need IT + ML to both succeed.
  • Insight generation (Analytics):
    • Measure data and reflect the product to tune it.
    • Redefine success.
  • Tuning: Automated feedback that is faster than many humans.

Inclusive ML: fairness of ML (Module 4)

  • Interactive bias: user interactions may inclined to interact with one over the other.
  • Latent bias: More male scientists than female ones during history

  • Selection bias: do you sure data represent everyone?

  • Confusion matrix: False Positive v.s. False negative
    • Select the appropriate metric according to the problem (Avoid FP or avoid FN is more important?)
  • Equality of oppurtunity: equal True positive rate for each groups.
  • A data visualization tool: Facets

Python notebooks in the Cloud

  • Serverless service:
    • Jupyter notebooks: DataLab.
    • BigQuery: aggregate big dataset using SQL.
  • Google pre-trained APIs:
    • vision
    • intelligence (for image/video)
    • speech
    • translation
    • natural-language

Thoughts

  • It is cool that google shared its procedure of transforming to a AI-first company by going through the five steps.