18 Followers
84 Following
103 Posts

Marin Benčević, Ph.D.



Improving the ways we train and evaluate deep learning models in medicine.



Postdoc @ FERIT, J. J. Strossmayer University of Osijek, Croatia

Going forward, preprints will form the basis of HHMI’s assessments of its researchers’ work.💪

Starting January 1,2026, HHMI will require its scientists to post their research articles as preprints under an open license that allows unrestricted reuse (CC-BY-4.0).

This new policy aligns closely with the response provided recently by ASAPbio to an NIH request for information.
We elaborate on the new HHMI policy in our recent blog post 👉
https://asapbio.org/hhmi-enacts-immediate-access-to-research-policy-for-its-scientists/

HHMI Enacts “Immediate Access To Research” Policy For Its Scientists – ASAPbio

The Howard Hughes Medical Institute (HHMI) is requiring its scientists to post their research articles as preprints under an open license that allows

PL - ASAPbio
@yoginho With respect, the term agent has been used in AI for decades, way before LLMs. And rational agent has been used since the 19th century and spans philosophy, economics, mathematics and other fields. I would not so easily make the leap that the term “agent” implies “human-like agency” when there is so much history of the term where it used to simply mean “an entity acting based on stimuli”.
I like to exit my vscode tunnels the same way I exit all of my social obligations.
@kjhealy also Section 5 of the YOLOv3 paper in object detection (a real work incredibly widely used): https://arxiv.org/pdf/1804.02767
@sobol @notjustbikes I don’t think this is unique, many cities have 4 car lanes with tram stops in the middle: https://maps.app.goo.gl/boHYZyqX3CD2kDj66
Bevor Sie zu Google Maps weitergehen

@cnlohr I fully skipped over this toot because I thought it was an ad. I then remembered that Mastodon has no ads and did a double take. 10/10 pattern matching by my brain.

Excellent news! We got a small amount of funding to hire two people half time to help put together a Horizon Europe proposal for Safeguarding Research and Culture [1].

The idea is to develop the SciOp catalogue software to make distributed archives part of the Fediverse, riding on BitTorrent, and get more institutions, libraries and archives involved. With a focus, in this case, on Europe.

If you are interested in preventing digital cultural heritage materials from falling through the memory hole, preventing marginalised and racialised groups from being erased from view, decentralised, federated and peer to peer technologies and have experience with putting together EU Horizon grant proposals, have the right to work in the EU or UK and would like a part time job starting yesterday for a couple of months, please get in touch!

[1] https://safeguar.de/
[2] https://sciop.net/

#FediHire

Safeguarding Research & Culture

"As researchers we often say 'we need the data'. Today, the data needs us." — Kathy Reid

Safeguarding Research & Culture
@pbloem I think part of it is just that what qualifies as publishable lags behind technology. 10 years ago you could run a popular NN architecture on a different public dataset and that would be publishable. Today, that wouldn’t even be a master thesis. I think the AI papers/results will just be a tool we use to check hypotheses/hunches while producing higher-level papers. In CS something that used to be a Nature paper eventually becomes a basic YouTube tutorial as technology evolves.
I don't like NumPy https://dynomight.net/numpy/
I don’t like NumPy

it’s too hard

DYNOMIGHT

As a computer scientist, I often wish I had more training in social science methods. This paper is a really good example of things we computer scientists forget to think about. It describes how to model the measurement process and how to measure unobservable abstract constructs such as race, gender, emotion (all common DL biometric applications!) etc.

https://arxiv.org/pdf/1912.05511