Bryor Snefjella
@b_snef
Assistant Professor interested in corpus linguistics, psycholinguistics, and social psychology. Latest paper: http://sciencedirect.com/science/articl…
Come join the CBI (cognition, behavior, and information) initiative at @asu_ssbs! Two positions available, one potentially at the associate level. Deadline Dec. 1st. apply.interfolio.com/114825
As part of my work on LLMs and analogy, the "but LLM training data is implausibly large" thing comes up a lot. Now the question is how performance is distributed across tasks for plausible training corpora.
Excited to share our new work with @martin_schrimpf, @zhang_yian, @sleepinyourhat, @NogaZaslavsky and @ev_fedorenko: "ANN language models align neurally and behaviorally with humans even after a developmentally realistic amount of training" 1/n biorxiv.org/content/10.110…
Oh man just wait until they get to best-worst scaling and CLMMs.
Excited to share our new work with @martin_schrimpf, @zhang_yian, @sleepinyourhat, @NogaZaslavsky and @ev_fedorenko: "ANN language models align neurally and behaviorally with humans even after a developmentally realistic amount of training" 1/n biorxiv.org/content/10.110…
Before everyone leaves twitter pls share with any promising UG students that my lab and the Cognition Behavior, & Information group at ASU are recruiting (funded) MS students for 2023! My lab: cognitivedatascience.org CBI: newcollege.asu.edu/cbi
Join a high-energy team in a growing School at ASU to help us transform science. @asu_ssbs is hiring TWO tenure-track / tenured faculty in Psychology or proximate field as part of a continuing expansion of our Cognition Behavior & Information initiative. apply.interfolio.com/114825
Do causal inference folks have anywhere they keep track of causal inference related job openings? Because I've got two in psychology at assistant/associate level: apply.interfolio.com/114825 @yudapearl @rlmcelreath
It was so good to be back home in Canada and see all my favorite wordnerds #mentallexicon2022
Heather Wild @gottabewild’s talk “Valence effects in the wild: Analysing word learning in language learning apps” was awarded “Best Talk” at #MentalLexicon2022 🎉 Congratulations Heather!
We need to keep asking under what conditions can cross validated error be taken as a valid indicator of generalization error? And do these conditions hold for where LLMs are getting deployed? When validation sets and deployment sets aren't from the same pop, CV is optimistic.
This overlooks an important aspect of the initial comment though, which is that LLM eval tasks (including probably more than half of BIG Bench) mis-represent what they are measuring; scaling ofc improves perf on those tasks but both human & ML metrics on such tasks are misleading
I'm just going to Osgood's The Measurement of Meaning books.google.com/books/about/Th… to heap of foundational stuff the PNAS paper missed.
Eye tracking folks take note! Will this unseat GAM(M)?
New preprint! "A Deep Learning Approach to Analyzing Continuous-Time Systems" (with William Schuler) arxiv.org/abs/2209.12128 We propose a way to use deep learning to *analyze* (not just imitate) complex continuous-time processes. 1/
As a Canadian postdoc in the US, I wonder if American postdocs are more training-oriented because of the lack of separate masters programs in the US. The American timelines and expectations for PhDs seem bonkers to me.
Just wanted to say in regards to this thread that @IbanDlank put his money where his mouth is and gave me a raise I asked for because I would be supporting a family during my postdoc with him.
The word “training” has been used to justify unfair pay, at least imo. Saying “salary aside, post-docs are trainees” is at best blissful ignorance about the power of that word. p.s. drawing the “training” line after post-doc vs. after PhD vs. after tenure is somewhat arbitrary
It does seem like a fail that the fantasy story made because of a conlang chose to copy real world accents for characters, instead of just... making up some new accents based on the original conlangs or a new conlang to inspire an accent. They had the budget. #lotr #RingsofPower
This paper is fighting the good fight for good measurements! A sad thing is how many old ratings datasets have their trial-level data lost to history, making updated versions scored with these models impossible. Guess someone has to recollect them all...
Finally out in Behavior Research Methods! Use mixed effects models to improve the quality of your inferences in norming studies 😀 1⃣2⃣3⃣4⃣5⃣ sad happy ☑️ doi.org/10.3758/s13428…
Finally out in Behavior Research Methods! Use mixed effects models to improve the quality of your inferences in norming studies 😀 1⃣2⃣3⃣4⃣5⃣ sad happy ☑️ doi.org/10.3758/s13428…
One thing I'm a little surprised by is 1) how useful and flexible #gamlss seems (like I can think of a bagillion models it can fit that no other frequentist software can, like doing lasso on predictors of dispersion in a mixed model) but 2) no one I know seems to use it. Why?