Adapting a Model You Did Not Train
Four Other Places to Put the Change
Last timeWhich Weights, How Wide, How Loud
Inserted layers, learned prompts, bias terms alone and a compressed frozen base are the other cheap methods, and each puts the change somewhere different at a different cost.
The thin-matrix construction is the one most teams use, but it is one member of a
family, and the others are worth knowing because two of them solve problems it
does not.
The one question behind all of them
Every method here follows the same recipe: freeze almost everything, pick a
small set of values to move, and arrange for those values to influence the
computation. What separates them is only where the chosen values sit.
Inserted layers
The earliest of these methods inserts a small module after existing parts of
each block, usually a squeeze down to a narrow width and an expansion back,
with the surrounding weights frozen. It trains cheaply and reaches good quality.
Its weakness is structural. An inserted layer is a new step in the computation
and there is no way to absorb it into an existing weight, so the extra
arithmetic and, more importantly, the extra sequential step are paid on every
request for the life of the deployment. In a service answering a large volume of
requests that permanent cost dwarfs any saving during training, which is exactly
why the thin-matrix construction displaced it.
The lesson stops here
6 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