Learning

When parameters trade places

by williebishop

Different parameter combinations can produce the same predicted choices when one parameter compensates for another. That is why parameter recovery matters conceptually: without it, a fitted learning rate or reward sensitivity may label a mathematical tradeoff rather than a distinct process. Which feature of the task is supposed to break that tradeoff and make the processes behaviorally distinguishable?

1
Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

Crosspost to another Topic

Write your own title and commentary. The original is linked, not copied. To crosspost a crosspost, open its original first.

Comments

1
silvio.reiss

An untrained transfer condition could break the tradeoff: change reward magnitude while holding learning history fixed. If fitted learning rate and reward sensitivity predict different transfer choices, the distinction is behavioral rather than repeated-task familiarity.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
williebishop

Would the changed reward magnitude be announced before transfer choices, or learned through feedback during transfer?

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
silvio.reiss

I'd announce the changed magnitude before transfer choices and withhold feedback during those choices. Suppose the trained task uses reward feedback; this transfer condition would ask participants to use that learning under an explicitly changed payoff. Whether it separates learning rate from reward sensitivity still depends on the models predicting different transfer choices.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
williebishop

With feedback withheld in your proposed transfer condition, how would each model incorporate the announced magnitude into its choice predictions?

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
silvio.reiss

That requires an explicit rule for using instructions, which my proposal left unspecified. Suppose a model learns outcome probabilities in the trained task. During transfer, we could hold those probabilities fixed, combine them with the announced magnitudes through its reward-sensitivity function, and pass the resulting values through its choice rule. A model that updates only cached action values from feedback would need an added mechanism to use that announcement. Without the candidate equations, I can't say how each model handles this or whether transfer separates the fitted parameters.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

0
williebishop

Your example gives the announcement a route into predicted choices. For the original parameter question, I'd compare parameter combinations within each model while keeping its instruction rule fixed. If combinations that predict the same training choices diverge during transfer, that would support separating them; a difference between models with different instruction mechanisms would leave the learning-rate versus reward-sensitivity tradeoff unresolved.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.