Deciding to Retrain
Last timeWhat Deserves an Alarm
Retraining is neither free nor safe, so the decision is a comparison between the measured decay and the cost, and a model retrained on a schedule nobody questions is its own failure mode.
What retraining actually costs
The quality has fallen four points. The obvious response is to retrain, and the
obvious response is sometimes wrong, so it is worth seeing what is actually
being spent.
| typical share of the tot | paid every time | gets cheaper with practi | |
|---|---|---|---|
| compute for the training | 5 | 1 | 1 |
| engineering time to asse | 25 | 1 | 1 |
| checking the new model b | 30 | 1 | 1 |
| deployment and rollback | 20 | 1 | 1 |
| a new set of unknown fai | 20 | 1 | 0 |
Against that sits the benefit, which is the recovered quality multiplied by
what quality is worth. A model handling two hundred thousand decisions a
quarter where each point of accuracy is worth four units of value recovers
twelve thousand units from a one and a half point improvement, and the whole
decision is whether that number exceeds the bill above.
The decay line
The decision becomes much easier once the labelled history exists, because the
history has a slope and the slope predicts a date.
The lesson stops here
5 more paragraphs to go
You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.
The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.
See the planThe contentsThis is the reading half
Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.
The contents