Ewa Dominowska, Engineering Manager, Facebook at MLconf SEA - 5/01/15
-
Upload
sessionsevents -
Category
Technology
-
view
164 -
download
2
Transcript of Ewa Dominowska, Engineering Manager, Facebook at MLconf SEA - 5/01/15
Managing Machine Learning Projects in Industry
Ewa Dominowska
Facebook, Engineering Manager
Agenda
• Building a Team
• Selecting and Framing a Problem
• Problem Solving Approach
• Evaluating Solutions
• Delivering Impact
Building a Team
• Engineer + ML Expert + Statistician + IR Expert
• Domain expertise
• Academic vs. industry experience
• Research + engineering + experimentation = applied research
• Investing (domain) vs. outsourcing (science)
Selecting a Lead
ML Expert
Engineer Manager
1 8
8
8
Organizational Structure
• Centralized Research Team
• Centralized Applied Research Team
• Embedded Researchers
• Team Members
• Academia
• Conferences, competitions, data releases, benchmarks
MSR, Facebook AI Lab
LiveLabs, FB Applied ML
Source: Bonkers World
Motivating
• Intellectual challenge
• Creative work
• Autonomy
• Purpose
• Mastery
• Recognition
• Publishing
• ConferencesSource: Motivationhacker
Selecting and Framing a Problem
Start with a business problem
Break down the problem
Understand the impact
Find the right data
Select an objective function
Build Models
Measure and Evaluate
ExperimentProductionalize
/ Scale
Problem Solving Approach
• Establish a baseline
• Check your assumptions
• Select a model\learning technique
• Select features
• Measure and evaluation
• Experiment
• Stability, scalability and robustnessSource: Sheldoncomics
Evaluating Solutions
• Defining the right metrics
• Offline evaluation
• A|B testing
• Meaningful vs. representative
• Representativeness and stability of results
• Offline vs. online metrics
• How to split traffic
• user, request, budget effects
• How long to run a test
• statistical significance, power, seasonality, novelty
• Calibration
• Model interactions
• Residue effects from previous experiments
Experimentation – Practical Lessons
Delivering Impact
• Plan for valuable failure
• Measure long term/steady state effects
• Engineering improvements
• Re-use of components, tools, models and frameworks
• Durability and robustness
• Data, context changes
• Measurement, monitoring, experimentation
Thank you!
We are hiring at Facebook!