search

joearedmond / dist-shift

Application that auto-retrains when it detects a distribution shift in production data.
0 stars 0 forks source link

Proof-of-concept retrain system on simple function model #1

Open joearedmond opened 2 years ago

joearedmond commented 2 years ago

How much data is needed to learn on a shifted distribution?

joearedmond commented 2 years ago

This may be two issues:

  1. naively re-train (fine-tune) after each test example is fed to the API (online retraining), or
  2. re-train after N test examples are seen (semi-online batch retraining).