Discover and read the best of Twitter Threads about #Regularization
Most recents (2)
(Thread) Here is a short summary of "Rules of ML" by Martin Zinkevich (developers.google.com/machine-learni…). The #MachineLearning life cycle should consist of three phases:
Phase 0
- Launch the product without #MachineLearning . Collect #Data
Phase 1
- Building the first basic/transparent #model
- Make the #infrastructure right
- Monitor the right set of #metrics
- System should get a boost with a simple model itself
- Launch the product without #MachineLearning . Collect #Data
Phase 1
- Building the first basic/transparent #model
- Make the #infrastructure right
- Monitor the right set of #metrics
- System should get a boost with a simple model itself
Phase 2
- More #FeatureEngineering on top of simple model (Low hanging fruits). #Regularization.
- Multiple launches with new features
- Metrics should still be rising with each launch
- More #FeatureEngineering on top of simple model (Low hanging fruits). #Regularization.
- Multiple launches with new features
- Metrics should still be rising with each launch
I cannot remember who and where introduced this sort of derivative regularization in #quantumfieldtheory of devergent integrals like: 
The #regularization was like assuming that the integral is not well defined but its derivative(s) are regular. Then:
It as the very nice property of giving a result compatible with dimensional regularization:
