Bespoke: Models are highly variable and play a variety of useful roles in scientific investigation and inference

Broke: all models are wrong but some are useful Woke: models are useful because they are wrong

ya agree. tbh it was more that i did not think you’d be a classical models guy, and i think the reference class of models is a sneaky but v important part of navigating this space, so the negative prediction error btwn my expectation & the post drove me to seek the relevant info lol

ahh gotcha okay i read too much into the initial post, you were saying that a criticism of classical models isn’t solved by ANNs either

perhaps more directly: what features of classical models make them appropriately/sufficiently(?) “brain-y” to you? fwiw i am asking bc i like to know where ppl fall on spectrums of takes lol, not bc i want to argue 🥲

im really curious what you have in mind as a classical model here, and why you think that classical models in general are the relevant contrast case. i’m generally sympathetic to this line of reasoning, but also see ANNs used in several ways that map to multiple diff construals of “brain like”

thank you!!!! it’s been so fun thinking with you about this for the past several years 🥲

For the first time in a while, a philosophy professor has been awarded one of the MacArthur genius grants...

big congrats to my mentor @felipedebrigard.bsky.social for being recognized for his incredible work on memory & forgiveness! also props to his student @gabrielafm.bsky.social who is leading this research and to lab manager @kmiceli.bsky.social

additionally, our model predicts that reliance on expectations ebbs & flows throughout a single decision: increasing at the moments in time when sensory evidence is uncertain, and decreasing as sensory uncertainty diminishes 〰️ this makes sense but has been difficult to test in lab---stay tuned!

critically, the findings we synthesize *cannot* be explained (or predicted!) by existing normative models. on our model, however, they emerge naturally from the temporally-extended, probabilistic nature of memory retrieval + optimal integration of that signal with incoming evidence from vision

we show that our model synthesizes neural & behavioral findings across four (!) experiments w/ diff manipulations of expectations (x&y axes below). also, we show that one of its key latent quantities---a time-varying weight on expectations--matches frontal theta oscillations in humans 🧠📈📉