New methods usually start from a specific, concrete limitation.
Read enough to be certain the idea has not already been published.
Prototype on the smallest problem that can show the effect.
Compare fairly against strong baselines, tuned as carefully as your method.
Most ideas fail, and the failures are where the understanding comes from.
New methods usually begin with a specific and concrete limitation.
This approach fails on long sequences is a starting point.
I want to invent something novel is not.
Read enough to be sure the idea has not already been published.
It usually has, in some form, and finding that is a useful result in itself.
The related work you find will sharpen the idea even when it does not kill it.
Prototype on the smallest problem that can show the effect.
If it cannot help on a toy problem it will not help on a large one.
Small problems iterate in minutes rather than days, which is where progress comes from.
Compare fairly against strong baselines, tuned as carefully as your own method.
The temptation to under-tune the baseline is enormous and it invalidates the work.
Run ablations to find out which part of your idea is doing the work.
Be prepared for it to be the boring part.
Most ideas fail, and the failures are where the understanding comes from.
Write down why it failed, since that is the part nobody else will publish.
One question on this card. Every option explains itself once you pick, right or wrong.
11 concepts in this module, then a graded test with every answer explained.