Acabamos de publicar nuestro trabajo!
🧠🤖
Los modelos de lenguaje superan a expertos humanos en predicción de resultados en neurociencia
[Nature Human Behaviour con @profdata.bsky.social, @ken-lxl.bsky.social, BrainGPT.org y muchos mas!] 1/9
Thanks so much to @rockberta.bsky.social, Yuri Bizzoni, and Kenneth Enevoldsen for inviting me to visit! The Interacting Minds Center is a really cool place: interactingminds.au.dk/events/singl...
Can we identify neuropsychiatric conditions from speech? Following papers on generalizability issues, @rockberta.bsky.social @giada.bsky.social Kritika Maheshwari and me now present an Ethics-Centered Approach to Research on LLM-Based Inference of Psychiatric Conditions: arxiv.org/abs/2409.15323 1/n
How has European identity changed over time? Did some fun comp text analysis work w/ @katharinalawall.bsky.social, @mtsakiris.bsky.social & Laura Cram, looking at the language (topic & style) of the European Commission's Twitter communication over time 👇
link.springer.com/article/10.1...
“This pain” or “that pain”? Can simple lexical choices be used to infer maladaptive mental states such as depression? Our new work hindawi.com/journals/da/... based on the Demonstrative Choice Task (DCT) suggests so. By Line Kruse with @rockberta.bsky.social and me. Thread in X
x.com/LineKruse4/s...
Big-team science for the win! It's been great fun to be involved in this project! 🧠 🤖 Detailed thread and link to preprint below 👇
Hey BlueSky! As a little weekend project, I have turned some of my #DataSci lecture notes into a blog post, now available on Towards Data Science 🥳 It's a little narrative primer on basic properties & common misconceptions on R-squared as evaluation metric for regression. Check it out!
Portrait of Alexei Navalny
Deeply disturbed and saddened by news of the death of Alexei Navalny.
Putin fears nothing more than dissent from his own people.
A grim reminder of what Putin and his regime are all about.
Let's unite in our fight to safeguard the freedom and safety of those who dare to stand up against autocracy
A close-up of Alexei Navalny posing for a portrait in Moscow in 2013. Photo by Sergey Ponomarev for The New York Times.
Breaking News: Aleksei Navalny, the most outspoken domestic critic of President Vladimir Putin, has died in prison, Russian state media said. nyti.ms/3wlPyvJ
se esiste un bluesky italico questo è un post d’amore assoluto per patty pravo, “pensiero stupendo” unica legittima utopia sociopolitica del secolo scorso
main points: interpretation of R^2 as % var explained (& equivalence w/ correlation^2) only applies to linear models fitted w/ least squares & in-sample. Out-of-sample & beyond linear models: there is no lower bound, models can "hallucinate" variance and R^2 is a comparative metric (vs mean model)
taught a fun first DataSci lecture for MSc students this week, and it turns out that debunking "myths" on properties (e.g., bounds) and meaning (e.g., % variance explained?!?) of R-squared is a great lead-up for a stats-to-prediction mindset shift. Hope to have time for a long-form post later on!
We have an opening for an associate/full prof in #NLProc, #cogsci, #digitalhumanities or adjacent areas in Aarhus. Vibrant research environment w/ lots of opportunities for collaboration, within and beyond disciplines. Please repost, and hit me up if you want to chat about Aarhus and the group. ❤️
Discussion on generalization / limitations / next steps also included in the preprint. Check out both preprint and package, and reach out if you have feedback! 😎
💡 Finally, preview of the results: 1) diversity is great for performance (# and originality or words named) if turn-taking is flexible; 2) turn-taking styles & diversity modulate search dynamics (e.g., exploration/exploitation); 3) individual flexibility interacts with diversity.
Secondly, the mechanics are fairly simple. Agents have word2vec models as semantic memories & take turns in producing word associations based on distance. Diversity is induced by adding noise to the w2v model. Turn-taking is modulated by defining strict or more flexible rules
👍
✍️ TL;DRs on rationale, methods, and main findings. Firstly, rationale. Manipulating mental representation and interaction structures "in vivo" is hard. W/ ABS, we can control individual, pair, and interaction-level factors, and assess emergent effects on performance and behavior.
The preprint focuses on two things: a) we introduce and present the paradigm, which we share as a 🐍 package (github.com/rbroc/simcat); b) we use it to study the effects of diversity & social interaction (turn-taking) on performance & behavior in a joint verbal fluency task 🧐
How do we search through mental spaces together? When are groups better than individuals? And how do we study this computationally? 🤔 In a new preprint, @kristian_tylen & I address these questions through a new agent-based simulation paradigm 🤖 💬 🤖: osf.io/preprints/ps...
How predictable is neuroscience? Can LLMs outperform humans? Please participate in the BrainGPT.org survey to help us find out. You choose between two versions of a neuro abstract: the original vs. one with altered results. Which is which? research.sc/participant/... 1/2
While FOMOing over #EMNLP and prepping last lecture of the term (efficient LLM training/fine-tuning), I made a quick viz to explain model distillation. Up for grabs if you ever get frustrated over the lack of good figures to explain this.
New paper! We benchmarked lots of text & speech models on lang-based inference of neuropsych conditions, found that, while binary classification performance is stellar, multiclass is generally low, and discussed why this may be so. Spoiler: it's not the models, but perhaps diagnoses as a constructs👇
looks cool! Thanks a lot for the pointer!
prepping a lecture on early transformers for next week, and trying to make my review of the BERTology more entertaining. any pointers to particularly *funny* BERTology papers (e.g., weird instances of model failure, or oddly specific knowledge BERT acquires through pretraining)? #NLProc
"Repeat after me? Both autistic & neurotypical children commonly align their language with that of their caregivers" in which we critically assess alignment and echolalia, out now in Cognitive Science: dx.doi.org/10.1111/cogs... w @ethanweed.bsky.social, @rockberta.bsky.social &al. A thread: 1/n
The Interacting Minds Centre is still Bluesky-less, but let’s see if we can do change that over the coming days :)
Stoked to have Giada Pistilli (@giada.bsky.social) visit us at the Interacting Minds Centre next week! Giada’s work navigates uncharted waters at the intersection of AI, ethics and policy-making with a fascinating combination of pragmatism and conceptual rigor. Talk will be hybrid - see link below!
grad students in need of a whole-brain fNIRS datasets (2+ tasks) to benchmark a CNN classifier they've developed. I have suggested openfnirs.org/data/ and tufts-hci-lab.github.io/code_and_dat.... Is there any other obvious resource I am missing and/or anyone owning such data and willing to share?
poster of the workshop
So excited about this hackathon-like workshop that'll take place in Leiden in Jan. We'll work in groups to develop & computationally implement cognitive models of complex behaviors (turn-taking, reading, word acquisition, etc.). Registration still open: lorentzcenter.nl/cognitive-mo...
looks brilliant, thank you!