-
Business Goals
- Advise users on interior design choices
- Decode the essence of style.
- user experience testing
-
Data Engineering Goals
- Build initial data pipeline
- Figure out how to integrate model with application.
- 1. Launch product with heuristics ASAP.
-
Incorporate ML into the product
- monitor performance as a function of "freshness"
- monitor feature coverage, watch for silent failures
-
Test infrastructure independently of ML
- Data Ingestion
- Exporting trained models
- unit testing
- system testing
- Develop visualzation/debugging tools
- Configure automated logging for online processes
-
Data Science Goals
- 1. Design and implement heuristics and metrics.
- Define what makes a system "good" and "bad"
- Consolidate high quality corpuses for ML
-
2. Shift from heuristics to simple machine learning models.
-
Define Objective
- don't overthink it
-
make ML objective simple. easy to measure, and a proxy for the "true" metric
- To avoid:
- Is the user happy using the product?
- Is the user satisfied with the experience?
- Is the product improving the user’s overall wellbeing?
- Choose simple features
-
Turn Heuristics into features.
- preprocess with the heuristic
- directly create a feature
- mine raw inputs of heuristic
- modify the label
-
start with an interpretable model
- debugging is easier
- serves as a baseline
- manage feedback loops
- decision matters more than the likelihood of the model
- Postpone spam considerations, focus on data posted in good-faith.
- ensure that model learns reasonable weights
- ensure features reach the model correctly
- automate "sanity-checks" before deploying models
-
3. Iterative improvement of simple ML
- Measure first, optimize second
- Featurize model errors
-
Move to more complex models that have been pioneered in the open source community
- Transfer Learning
- Finetuning
- end-to-end training
- ensembles