Nikolas Adaloglou
Düsseldorf, North Rhine-Westphalia, Germany
7K followers
500+ connections
7K followers
500+ connections
Activity
7K followers
-
Nikolas Adaloglou reposted thisDiffusion model can only use fixed Gaussian noise? Not anymore, our CVPR paper has been extended to arbitrary noise patterns while preserving the modularity and efficiency of EDM-style diffusion. "Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models", CVPR, 2026 https://lnkd.in/ghpcZCa3
-
Nikolas Adaloglou reposted thisNikolas Adaloglou reposted thisToo many REPA / RAE / representation alignment papers lately? I was lost too, so I wrote a blog post that organizes the space into phases and zooms in on what actually matters for general/molecular ML. Curious what folks think - link below! 📎 Blog: https://lnkd.in/gmb6STBJThe unification of representation learning and generative modellingThe unification of representation learning and generative modelling
-
Nikolas Adaloglou shared thisThis week, I defended my PhD thesis, "Designing representation learning applications for unlabelled image data," with a magna cum laude grade! I am looking forward to what's coming ahead for 2026! The AI landscape has been significantly different over the last 3-4 years. And I hope I can be part of something great quite soon!!! I am primarily interested in industry research-related positions in Europe. Happy to chat about work opportunities in a DM. Notably, one of our works has received the best paper award at #BMVC 2025 ("Guiding a diffusion model with itself using sliding windows"), and I will also present our latest work on out-of-distribution detection, ClusterMine, at #WACV 2026 in March. A big thank you from the bottom of my heart to all my friends and family who showed up, as well as my supervisor, Prof. Markus Kollmann, and the amazing committee, for their constructive feedback and questions.
-
Nikolas Adaloglou reposted thisNikolas Adaloglou reposted thisMistral just dropped the Ministral 3 tech report, and it's a great example of how you don't need massive compute to build competitive small models. they trained their 3B/8B/14B models on only 1-3 trillion tokens. the trick? smart pruning + distillation the approach is pretty clean: start with their 24B model, progressively prune it down to smaller sizes, then use the original model as a teacher to recover performance through distillation. each smaller model is initialized from the pruned weights of the previous one, but they all learn from the same 24B instruct teacher the pruning itself is interesting. they prune depth, hidden dim and ffn dim, each with a different method. for layers they look at how much each layer transforms its input (output/input norm ratio). for hidden dimensions they use PCA to find the important directions since features aren't axis-aligned, then prune the low variance ones. for FFN they look at the gated activation score since in SwiGLU a high value can still be killed by a low gate they have some cool ablations too. using an instruct model as teacher works better than base for STEM tasks. and they show a capacity gap effect where a bigger teacher can actually hurt, Medium 3 (larger model but size is not public) as teacher performed worse than Small 3.1 for pretraining but better for post-training solid paper with good ablations, bravo Mistral AI!
-
Nikolas Adaloglou reposted thisNikolas Adaloglou reposted thisText-to-image diffusion transformer models learn to align text and image representations as a byproduct of their conditional denoising task. By taking the dot product between the text and image representations of a DiT model (like Flux 2), you can create rich saliency maps. Our method ConceptAttention creates rich saliency maps of text concepts present in generated images and videos. I implemented it for the recent Flux2 model which has a higher maximum latent resolution (128x128 I believe) and found it still has the same properties as on earlier multi-modal diffusion transformer models. It requires no additional training, only repurposing existing parameters. Paper: https://lnkd.in/eFjwbicN Code: https://lnkd.in/e_b-ieaX
-
Nikolas Adaloglou reposted thisNikolas Adaloglou reposted thisWe are proud to announce a strategic partnership with the Government of the Hellenic Republic! This cooperation is based on our Mistral AI for Citizens framework, and will allow the following initiatives: 🇬🇷 Greek models: leveraging Greece’s open source resources and academic expertise to build the best Greek-speaking models 🔍 Joint research: exploring ways for developing innovative AI products with a strong focus on Greece’s maritime sector ⚡ Digital transformation: enhancing the efficiency of the Government’s processes and services 🧠 Talents: building a Talent Accelerator Program between National Technical University of Athens and Mistral AI We thank the Greek government for its trust, and are happy to keep pushing for Europe’s strategic autonomy. Stay tuned for what’s next!
-
Nikolas Adaloglou posted this✈️ 🚀 Heading to #EurIPS 2025 in Copenhagen tomorrow, full of energy to learn more about the latest advancements! After a spectacular time in #BMVC 2025, I am looking forward to meeting and discussing future job opportunities and collaborations. If you're hiring in Europe for computer vision and research engineering/scientist roles, I’d love to meet or grab a coffee ☕️. See you in Copenhagen 🙌 Have a fab. week!
-
Nikolas Adaloglou shared thisGoing to #BMVC 2025 next week in Sheffield? Let us meet! We will be presenting our latest work "Guiding a diffusion model with itself using sliding windows" on Monday as an oral presentation. If you are looking for fresh PhD graduates for your team in areas including but not limited to generative learning, vision-language models, diffusion models, visual representation learning, out-of-distribution detection, large scale clustering, send me a DM. Links in the comments!
-
Nikolas Adaloglou reposted thisNikolas Adaloglou reposted thisMost OOD detection methods still rely on predefined in-distribution labels. That creates problems: labels can be incomplete, labor-intensive to create, or poorly suited to open-world settings. ClusterMine takes a different approach that avoids ground-truth labels entirely. Developed by Nikolas Adaloglou, Diana Petrusheva, Mohamed Asker, Felix Michels, and Markus Kollmann from the Heinrich-Heine University of Düsseldorf, it combines visual clustering of CLIP features with zero-shot image-text scoring to identify relevant concepts from large text corpora. This lets the model define "in-distribution" from the data itself rather than relying on static label sets 📊 Across datasets like NINCOv2, SUN, and SSB, ClusterMine reached strong AUROC scores (95.86 on NINCOv2) and performed well under covariate shift, including on ImageNet-C and ImageNet-R. That said, this doesn't make human-labeled data obsolete. ClusterMine is designed for monitoring and boundary detection, not core supervised training. And it still depends on models like CLIP, which were trained on human-curated datasets. What it does change is how we define distributions. When task-specific labels aren’t reliable or drift over time, broad text corpora like WordNet or ImageNet-21K offer a more flexible foundation. ClusterMine builds on that to create a practical alternative to fixed-label supervision in safety-critical and evolving environments. For teams working on open-world systems, perception, or model reliability, this is worth a close look. Leaving the link to the paper below. ⤵️ #OutofDistribution #SelfSupervised #ConceptMining #VisualMonitoring #MLOps
Nikolas Adaloglou commented on a post
2mo
Congrats!
Nikolas Adaloglou replied to a comment
7mo
Timo Dickscheid Thank you from the bottom of my heart Timo!!!! I really appreciate your kind words!
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked thisI recently put together a small GitHub repository aimed at Machine Learning practitioners interested in the Ad Tech space: https://lnkd.in/diBXQXid Bridging the gap between general ML and ad-tech-specific concepts can be tricky, so I gathered some resources that I wish I had when I was first starting out . It's just a starting point, but I hope it can be useful to anyone looking to jump into this field 😵💫. I'm always looking to learn more, so feedback, suggestions, and contributions are very welcome! #MachineLearning #AdTechGitHub - nimaous/Ad-Tech-For-ML-Practitioners-: Ad Tech for ML/RL Practitioners: A hands-on introduction to advertising technologyGitHub - nimaous/Ad-Tech-For-ML-Practitioners-: Ad Tech for ML/RL Practitioners: A hands-on introduction to advertising technology
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked thisGreeks in AI has established itself as one of the leading meeting points for the Greek AI community, bringing together researchers, innovators, entrepreneurs and industry leaders from Greece and abroad. This year, Information Technologies Institute (ITI) participated with a large team of researchers, contributing to technical talks, workshops and poster presentations highlighting the breadth and quality of our AI research. Our work spans multimodal AI, vision-language models, trustworthy and fair AI, AI-generated content detection, scientific information retrieval, industrial AI and multimodal video understanding, reflecting our commitment to advancing AI technologies with scientific excellence and real-world impact. Congratulations to the Greeks in AI organizers for another excellent event that continues to strengthen the Greek AI ecosystem and foster collaborations between academia, research and industry. Posters presented by the ITI team: M. Krestenitis, C. Tzelepis, K. Ioannidis, S. Vrochidis, I. Kompatsiaris, G. Tzimiropoulos, S. Gong, I. Patras, CycleCap: Improving VLMs Captioning Performance via Self-Supervised Cycle Consistency Fine-Tuning (arXiv:2603.18282) N. M. Militsis, A. Toumpas, I. Koulalis, K. Ioannidis, S. Vrochidis, EviInspect: Evidence-Grounded Annotation and Evaluation for Safety-Critical Industrial Inspection (CVPR 2026 Workshop) M. Mylonas, C. Zerva, E. Apostolidis, V. Mezaris, SD-MVSum: Script-Driven Multimodal Video Summarization Method and Datasets (arXiv; earlier version presented at ACM Multimedia 2025) Y. Sarridis, C. Koutlis, S. Papadopoulos, C. Diou, MAVias: Mitigate Any Visual Bias (ICCV 2025) J. Bakagianni, S. Papadopoulos, MeVer at CheckThat! 2026: Cluster-Aware Hard-Negative Mining for Multilingual Scientific Source Retrieval (CLEF 2026) T. Pantsios, D. Karageorgiou, C. Koutlis, G. Karantaidis, O. Papadopoulou, S. Papadopoulos, Automated In-the-Wild Data Collection for Continual AI Generated Image Detection (ICMR MAD 2026)
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked this📢 New paper shoutout from the time in Boston! SynIB: An Informational Bottleneck for Maximizing Synergy in Multimodal Learning. If you've trained multimodal models, you've probably seen this: most of the time they don't really use both modalities. They lean on whichever one is easiest to fit and ignore the rest. Great average accuracy, but they fall apart on the examples that genuinely need both. We found the cause surprising: the cross-modal ("synergistic") signal isn't out-competed as thought, it's just rare, so the model memorizes it instead of learning it. An optimization problem, not a capacity one. SynIB fixes it with one idea: during training, mask a modality and penalize the model if it stays confident. Confidence that survives masking = a unimodal shortcut, so we push back on it. A pure loss term, drops into any fusion model. We also release CREMA-D-Irony, a benchmark where you can dial cross-modal synergy up or down. 📄 Paper: arxiv.org/abs/2606.09853 💻 Code: github.com/kkontras/SynIB 🤗 Dataset: https://lnkd.in/eRSgngiR Huge thanks to my co-authors across KU Leuven and MIT: Teodora Popordanoska, Thomas Strypsteen, Christos Chatzichristos, Matthew Blaschko, Maarten De Vos, and Paul Liang. #MultimodalLearning #MachineLearning #DeepLearning #AI #Research
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked this🚀 Some personal news: as of this month, I have been appointed Professor of Physical AI at the University of Amsterdam. Nearly two decades ago I started with a simple question: how do machines see? That question kept evolving. Computer vision taught me that recognition is not enough. In 2016 we introduced Siamese tracking — following objects through time. Time turned out to be the real subject. So the work shifted to the fundamentals of learning dynamics: how neural networks represent change, motion, and the passage of time. Dynamics led to causality — cause and effect, interventions, what happens if. And causality led to mechanisms: neural networks that carry physical structure and explicit state inside their architecture, not just weights. And now it all converges. Understanding, modelling, and designing the physical world. World models you can build robots on. Real-to-sim digital twins. Robot learning that generalizes because it respects physics, not despite it. That convergence has a name: Physical AI. I am deeply honoured that the University of Amsterdam has made it a professorship. My honest take: the timing is no accident. Foundation models gave us scale. What Physical AI needs next is structure — mechanisms, causal reasoning, explicit state. Actions are not text. All that is the research programme, and it starts now. None of this happened without support. My gratitude to my mentors Arnold WM Smeulders, Cees Snoek, and Max Welling and my colleagues at the Informatics Institute at University of Amsterdam, to European Research Council (ERC) and NWO (Dutch Research Council), and my partners at Toyota Motor Corporation /Toyota Technological Institute at Chicago (TTIC) , Qualcomm , and Elekta — and above all to my PhD students and postdocs, past and present. Your energy and contribution opened the way to this professorship. To the next chapter: robots that understand the world they act in. Safely, reliably, by construction. #PhysicalAI #Professor #UniversityOfAmsterdam #Robotics #RobotLearning #WorldModels #CausalAI #MachineLearning #ComputerVision #AI #UvA
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked thisJust got back from MIDL Conference 2026, held this year in Taipei, Taiwan and it delivered on two fronts at once. On the science side: strong attendance, high-quality papers, and the kind of hallway discussions that make a conference worth the flight. I had the opportunity to present our work, Context-Aware Patch Representations for Multiple Instance Learning. In short, if you work with Multiple Instance Learning, and pathology in particular, we introduce a new correlation operator that acts in the pre-aggregation space, which, 1. Soft-clusters the N patch embeddings (~10k) into a handful of M cluster centroids (~4–6). 2. Correlates those few centroids via self-attention. 3. Recomputes the N patch embeddings as a linear combination of the correlated centroids. Paired with a simple mean aggregator, this matches or leads far more complex MIL heads on discrimination and calibration, while cutting parameters by 50–93%, and it plugs into any MIL aggregator, making it cheaper to run too. On the other side: Taipei itself. There was a whole city to take in, centuries-old temples, the skyline anchored by Taipei 101, and sunsets over the hills that were hard to look away from. A rare combination of a strong research community and a stunning place to experience it in. A big thank you to the MIDL organizing committee for putting together such a wonderful event. P.S. This work was supported by ARCHIMEDES Research Unit on AI, Data Science, and Algorithms
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked thisImage foundation models never stop surprising! 😮 And this time, it's not the usual big tech players. A few days ago, Robbyant released LingBot-Vision, a new family of vision encoders built natively for dense spatial perception. Available in ViT-g (1.1B) alongside distilled ViT-L, ViT-B, and ViT-S sizes, and with a fully permissive commercial license, these models introduce a new pretraining task focused on fine-grained geometry. Altering the DINOv3 recipe, the teacher identifies patches containing structural boundaries and masks them for the student in the iBOT loss instead of a random selection. For these tokens, the student must predict both the latent feature and the edge properties. The spatial precision pays off immediately. Trained on a dataset of comparable size to DINOv2 (161M images), the ViT-g model outperforms the 7B DINOv3 on tasks like depth estimation, and the distilled models score favorably all across the board. Downstream, these backbones power LingBot-Depth 2.0, a depth completion model for robotics and perception applications. Trained on a whopping 150M samples, the model can accurately reconstruct challenging structures like glass, mirrors, and transparent objects. Exciting times ahead as we wait for the response from known labs. One thing is for sure: Robbyant is shipping a ton lately, and once again, mere scaling is not always the answer! Looks like new SSL image models will require fresh ideas and harder objectives to succeed. All links in the comments, enjoy! ⏬
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked thisVery excited to share that our paper "ExPLoRe: Expert Patch-Level Loss Routing for Multi-Objective Masked Image Modeling" has been accepted at #ECCV2026. It is the last paper from my PhD at the University of Tennessee, Knoxville, and I was fortunate to work on it with Maofeng Tang and my advisor Hairong Qi. Multi-objective masked image modeling often combines several learning signals (for example token distillation, CLS alignment, pixel reconstruction). Most methods weight these with global scalars, one number per objective. ExPLoRe gives each patch its own loss weight, learned by the router. A few takeaways on the method: • We reuse Soft-MoE dispatch weights as per-patch loss coefficients. • Loss-coupling: the MoE router is implicitly trained by the same loss signals it distributes across patches. • Different experts learn to weight specific patches for each loss, distributing competing per-patch signals across the losses. The specialization is implicit: it emerges from training itself. Paper: https://lnkd.in/dhRy84pZ Code: https://lnkd.in/dCSCA_kv Blog-post: https://lnkd.in/d9Md6Wcw
-
Nikolas Adaloglou liked thisNikolas Adaloglou liked thisOn May 7th and 8th, we had the pleasure of attending the 1st Annual Meeting of ESAC - European Interdisciplinary Society of AI for Cancer Research and the 4th AI for Oncology and Cancer Research meeting. It was a wonderful opportunity to share science, ideas, and connect with both familiar and new colleagues. The two days were filled with inspiring discussions, diverse topics, and outstanding professionals from the field. We were also glad to present the preliminary results of our work on HRD prediction from H&E-stained whole-slide images in ovarian cancer patients. A big thank you to everyone who stopped by our poster and shared comments, questions, and suggestions. Your feedback was truly valuable and will help us move this work forward. Many thanks to the organisers for putting together such a rich and engaging event. We look forward to seeing you again next year!
Experience
Education
-
University of Düsseldorf
-
Title: Designing Representation Learning Applications for Unlabelled Image Data using Visual Foundational Models
-
-
-
-
Courses
-
Artificial Intelligence
-
-
Bioinformatics
-
-
Biostatistics
-
-
Embedded Microsystems
-
-
Medical Imaging
-
-
Medical Robotics
-
-
Object Oriented Programming
-
-
Pattern Recognition
-
Languages
-
English
Full professional proficiency
-
Greek
Native or bilingual proficiency
-
French
Limited working proficiency
-
Spanish
Limited working proficiency
-
German
Elementary proficiency
Recommendations received
1 person has recommended Nikolas
Join now to viewOther similar profiles
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More