I find this tendency of training #AI from your data in your PDF, emails, etc quite perturbing. The companies behind will claim that the data is used just to learn general abstract rules. The reality is that these models memorize parts of the data, and they might repeat it. This #memory is an emerging property. Actually, don't babies repeat and remember when they learn?
Remote
Miguel Colom
@mcolom@mastodon.social
mastodon 4.8.0-nightly.2026-10-06Researcher (PhD, HDR) at Centre Borelli, ENS Paris-Saclay. ICPR 2026 Reproducibility Chair. Professor at the "reproducible research track" of the ENS Paris-Saclay MVA master.
0 Followers
0 Following
17 Posts
Joined August 19, 2023
website:
github:
website ENS:
I'm at the STI-ENID conference in Bristol now. Amazing place, I feel like Harry Potter in here 
Our #research paper "Combi-CAM: A Novel Multi-Layer Approach for
Explainable Image Geolocalization" by D. Faget et al. will be presented at VISAPP 2026. It's on geolocation explainability and there's a closely related preprint here: @836918@www.techrxiv.org
If you're interested in improving code sharing in #health research, @NaudetFlorian@nerdculture.de has a open #postdoc position on evaluating code sharing in BMJ publications. More info here: https://www.bennett.ox.ac.uk/blog/2025/08/help-us-improve-code-sharing-in-health-research/
Going today's to the Software Heritage event at #UNESCO. Moreover, there's an exhibition around software that one can visit during coffee breaks, it seems. That's great! Not only the free coffee, but the software too :D
Tomorrow Jun 11 I'll be presenting the #SWHID (the Software Heritage Persistent Identifier) at #pidfest, an intrinsic identifier that points to permanently stored source #code, thus allow for #open #science fostering the scholarly ecosystem and much more! I'll briefly how the #IPOL journal integrates it in our publication workflow, as a use case. A joint work with @moranegg@mstdn.social, the SWHID WG, and the Software Heritage team. #swh #research
Uhm, "General Medicine" small, and "Homeopathy" very large. I wouldn't go to see this doctor!
In collaboration with CESBIO we're proposing an M2 #research internship at Centre Borelli (ENS Paris-Saclay) on the reconstruction of interferometric satellite images with deep learning techniques. If you're interested in this internship to follow up with a PhD (if you manage to obtain a scholarship!), please let me know. For candidates outside the EU, note than you'd need to obtain the authorization of the French Défense Nationale before applying. http://mcolom.perso.math.cnrs.fr/download/stages/Stage_CESBIO_2026_NMS.pdf
The recordings of the four sessions of "EUDTP 2025: EU Digital Technologies and Policy Conference" are now available on Youtube, https://www.youtube.com/playlist?list=PL2xhU17lY9ZH5DAGKV_GXk4VeMfP8-yLM
A copyright troll company is asking me for thousands of euros because they found a version of one of their images in my personal, non-profit, research-oriented website. We'll see how this ends...
Thanks Oleksandr Berezko from the UARN and co-organizers for the invitation to talk at the "Training course ‘Peer Review in the Era of Open and Reproducible Science: New Challenges and Opportunities". I was impressed by the amount of interested people attending this online course, congratulations! https://lpnu.ua/en/news/training-course-peer-review-era-open-and-reproducible-science
#research #course #reproducibility
Submissions to #ICPR 2026 (International Conf. on Pattern Recognition) are now open! We'll acknowledge the large effort made by authors towards open science and specifically on #reproducibility by granting a badge and giving a prize to the best work. Don't hesitate to submit and apply! See you in #Lyon ;)
On Oct. 30 at 2 PM Paris time (or 1 PM GMT) I'll be making an introduction to reproducibility and computational sciences in the series of seminars of AREN (African Reproducibility Network). If you're interested, don't hesitate to join! https://us06web.zoom.us/meeting/register/2kUnopebSy-DcY--Jljq_g#/registration
So unusual, empty train today...
I presented a tutorial on how to run code in our IPOL journal in the #acmrep conference. I could have used a pre-built example I knew would work flawlessly. However, I preferred to ask the audience for a real code, and make it work. Why? To see all the problems we could find in reality. Some of the trouble we had: an unusual language (R) in our community, public keys for accessing the code, dependency installation, and environment setup. I totally enjoy seeing my experiments failing miserably!