@mario-kddb.bsky.social
PhD Student on Foundation Models in Microscopy @HTW-Berlin
Robust Weight Imprinting: Insights from Neural Collapse and Proxy-Based Aggregation
New #TMLR-Paper-with-Video:
Robust Weight Imprinting: Insights from Neural Collapse and Proxy-Based Aggregation
Justus Westerhoff, Golzar Atefi, Mario Koddenbrock et al.
https://tmlr.infinite-conf.org/paper_pages/duU11BnQ3Y
#imprinting #imprint #neural
btw, the image above are the results from before using our approach. When optimizing SAM3 for this task, we reach this:
19.01.2026 09:05 — 👍 2 🔁 0 💬 0 📌 0
15.01.2026 18:23 — 👍 0 🔁 0 💬 0 📌 0
We developed a method to fine-tune synthetic microscopy images to closely match real ones + pixel-perfect segmentation masks included.
Bridging the gap between synthetic and real data when labels are scarce.
Preprint: datexis.github.io/SynthMT-proj...
ICML changed their reviewing system a tiny bit. blog.icml.cc/2026/01/08/w...
13.01.2026 07:54 — 👍 1 🔁 0 💬 1 📌 0I really like your package. works well :)
12.01.2026 22:06 — 👍 1 🔁 0 💬 0 📌 0congrats!
12.01.2026 22:05 — 👍 0 🔁 0 💬 0 📌 0I love the idea! Did it happen? I’m late to this thread :D
12.01.2026 21:20 — 👍 0 🔁 0 💬 0 📌 0If you’re interested in a systematic comparison of these two models (and some more), we conducted such a comparison on micrographs of microtubules.
datexis.github.io/SynthMT-proj...
I can confirm this on our micrographs of microtubules:
datexis.github.io/SynthMT-proj...
It worked surprisingly well even on our non-cell data. We tested it on micrographs of microtubules and it performed out of the box.
datexis.github.io/SynthMT-proj...
after evaluating both I would rather say CellSAM. But our data is very specific. Microtubule Micrographs: datexis.github.io/SynthMT-proj...
12.01.2026 21:06 — 👍 3 🔁 1 💬 1 📌 0Thank you for sharing the open-source code and open models. TARDIS is rather the exception in the domain. There are so many closed-source approaches or dependencies on proprietary software.
12.01.2026 21:03 — 👍 2 🔁 0 💬 0 📌 0thanks for sharing :)
12.01.2026 20:48 — 👍 1 🔁 0 💬 0 📌 0Our preprint is online ⬇️
12.01.2026 20:26 — 👍 0 🔁 0 💬 0 📌 0#SAM3 is getting a lot of attention, but how well does it work in specialized domains?
On microscopy data, it performs surprisingly well without any fine-tuning — just careful hyperparameter tuning. In fact, it reaches the same level of agreement as two human annotators.
#Preprint (under review)
🐤
11.01.2026 14:01 — 👍 0 🔁 0 💬 0 📌 0I used to think: "uv is fast, but is package installation really a bottleneck?"
Then I destroyed my conda env at 11pm while trying to fix a plot for a paper due tomorrow.
Now I get it. uv is incredible 😅
#Python #DevTools
First ML+BIO paper of my PhD journey submitted 🎉
Can't wait to receive the bioRxiv link to share our findings!
Thank you @overleaf.com for making collaboration so easy!
Just completed a paper with 5+ actively writing/reviewing authors where we worked through 300+ comments. Real-time editing with multiple people working on the same sections without a hitch.
Great piece of software :)
