The Honest State Of The Answer
Last timeThe Curve That Comes Back Down
Part of why very large models generalise is understood, part is conjecture with good evidence, and the remainder is genuinely open.
The previous lesson left a question rather than an answer. Past the threshold many
settings fit the training data exactly, the procedure picks a gentle one, and gentle
ones predict well. The first two parts of that sentence are solid. The third is where
the honest account gets uncomfortable, and this lesson is about separating what is
established from what is conjectured and what is simply unknown.
The choice nobody wrote down
Start with the part that is understood. When many settings achieve zero training
error, the training error cannot be what distinguishes them, so something else does,
and that something is the fitting procedure. It begins at small values and moves in
small steps, so it stops at the first adequate solution it reaches rather than the
largest, and in the simplest families it can be proved to arrive at the solution of
smallest size among all exact fits.
This is a penalty nobody added. It is not written into the objective, it cannot be
turned off, and it is doing a large part of the work usually credited to the
architecture. It also explains why the capacity lesson insisted that capacity belongs
to the procedure, since here the procedure is the only thing making a choice at all.
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