Theoretical Physicist and Deep Learning Scientist
Trieste, Italy
Joined November 2009
- Tweets27
- Following415
- Followers125
- Likes103
alessio ansuini retweeted
Excited to share that 2/2 papers from our Lab (LADE) were accepted to #NeurIPS2025 (one spotlight 🎉)
Great work from all the students and collaborators involved!
@AreaSciencePark @aleserra1998 @lorebasile @francescortu @lucrevaleriani @DiegoDoimo @ansuin @FrancescoLocat8
On the #BiodiversityDay2025 we are honored to support biodiversity conservation in Madagascar.
Thanks to @TWAS-@Sissa-Lincei fellowship,
@AreaSciencePark and the University of Antananarivo are using AI for monitoring wood exploitation in Madagascar.
alessio ansuini retweeted
📢Are you a scientist whose work can be applied to reach new heights in the study of pathogens? 👩🔬👨💻
The International Conference on Pandemic Preparedness is your unmissable chance to connect in a multidisciplinary network 🤝pathogen-ri.eu/conference/#o…
1/4 #prpconference2024
alessio ansuini retweeted
It was a great pleasure to deliver a LADE seminar at @AreaSciencePark in #Trieste! Big thanks to Francesca Cuturello and Alessio Ansuini @ansuin for having me here!
scientific-events.areascienc…
alessio ansuini retweeted
Please help me #RT this 2-years postdoc position #Bioinformatics / #DataScience at the crossing of single-cell #Genomics and #longreads sequencing. With @UniTrieste and @AreaSciencePark in Trieste! Fun work, people and science: more on #LinkedIn linkedin.com/jobs/view/37928…
alessio ansuini retweeted
Open expression of interest for two postdoc positions in #cancer #datascience at @AreaSciencePark and @UniTS on a joint project led by @gcaravagna and me.
Join us to develop cool computational methods for single-cell and long-read data!
More info👉 shorturl.at/lptNT
alessio ansuini retweeted
Great opportunity last week to teach at the PhD Summer School on self-supervised learning @DTU_compute
We discussed how we use geometry to understand LLMs representations with @ansuin
Thanks @Tweetteresearch for the invite
Slides coming soon😊
#beyondreductionism
Honoured to be part of this amazing project!
#DPCfam is a great method, and we are just scratching the surface of what it can do.
Stay tuned...
We increase protein family annotation of the Unified Human Gastrointestinal Proteome (UHGP) using DPCfam clustering.
Nice effort by our PhD Federico Barone @areasciencepark, @elena_tea @ansuin and great collaboration with Marco Punta @CosrLab
Preprint: bit.ly/42bV3HY
alessio ansuini retweeted
We increase protein family annotation of the Unified Human Gastrointestinal Proteome (UHGP) using DPCfam clustering.
Nice effort by our PhD Federico Barone @areasciencepark, @elena_tea @ansuin and great collaboration with Marco Punta @CosrLab
Preprint: bit.ly/42bV3HY
Less than 3 days left to join us!
If you are a data engineer and you have a crush for challenging scientific problems you can't lose this opportunity!
linkedin.com/jobs/view/35547…
We are building an AI-powered Large Research Infrastructure, and we need Data Engineers to make it real.
Deadline is approaching!!!
linkedin.com/jobs/view/35547…
alessio ansuini retweeted
See you at @workshopMLSB to discuss our work “The geometry of hidden representations of protein language models”.
Great collaboration with @lucrevaleriani, Francesca Cuturello, @ansuin at @AreaSciencePark!
biorxiv.org/content/10.1101/…
👇Summary below
#jobopportunity One research scholarship for machine learning with a new approach on Industrial datasets!
Deadline approaching: Thursday 10.03 (still 48 hours!)
areasciencepark.it/lavora-in…
Maybe a test would be to repeat the same analysis on artificially generated data with complex and controllable symmetries (we were close to obtaining one, on the other hand the artificial dataset we used in the paper was too small to do such tests properly.)
We cannot say that this really explains what is observed in ImageNet, that has much more complex features than MNIST (such as textures, pose, etc...), but we think this mechanism may be at work also there.
We thought that, in getting rid of the irrelevant feature,
the network produced representations, in the first layers,
in which the relevant features emerged more prominently,
and this resulted in an increase of their intrinsic dimension.
We trained the same architecture used for the original MNIST with the perturbed dataset, and, at the end of training, we noticed that in the first layers the dimensionality
grew and reached, in the middle, values similar to the corresponding values in the original MNIST.