Duc Tam Nguyen

🧠 Reflecting on: Why Graph Technology for AI Reasoning 🎙️ Speaking at DSC DACH was an incredible experience. It was a unique opportunity to dive deep into one of my favorite topics — how Graph Technology enhances AI, particularly through RAG — and engage in discussions with brilliant minds. We tackled key challenges like explainability, privacy, ethical responsibility, and knowledge management automation. 💡 Key takeaways: - Enhanced contextual understanding: Graphs create richer AI models through interconnected relationships. - Improved retrieval: Graph RAG allows AI systems to access diverse, private data sources, reducing biases and improving decision-making. - Privacy protection: By structuring and retrieving data securely, Graphs ensure sensitive information isn’t exposed. - Increased explainability: Graphs shift AI away from fully "black-box" models, offering clearer and interpretable insights. - Automated knowledge management: Graphs help integrate human domain-expertise with processes data, preserving institutional knowledge, enhancing decision-making, and driving innovation. 📖 For a deeper dive into these concepts, check out my article: Why Graph Technology for AI Reasoning (https://lnkd.in/dHHbq3NN) Looking forward to bringing more data stories to life! 🚀 #DSCDACH2024 #AI #Graphs #GenerativeAI #RAG #EthicalAI #Explainability #KnowledgeManagement

118

5 Kommentare

Munich AI Founders Night: Last Chance Alert!🚨 Tomorrow is the day! Join us at the AI Founders Night and be part of Munich’s thriving AI ecosystem! 🌟 📅 When? Wednesday, November 6th at 6 PM 📍 Where? TUM.ai Homebase, Unite Office, Rosenheimer Str. 116A, 7th floor 🚀 This is your final opportunity to secure your spot at an event designed to connect AI visionaries, innovators, and ecosystem partners. With a lineup of inspiring speakers and a room full of ambitious entrepreneurs, this is an evening you won’t want to miss! Speakers include: ✨ Sophie Defauw (CEO, AthleteIQ) TUM.ai alumna, started Uni at 13, became a professor at age 22, has been a guest lecturer in sport tech at UC Berkeley and won the Elon Musk 2021 Not-a-Boring competition. Furthermore she was a pro-athlete in windsurfing, top 5 in North America and made national art exhibitions. ✨ Florian Scherl (CEO, FAST AI Movies) and Philipp Gabriel (CTO) Founders and former AI E-Lab participants who met during the program’s final pitch. Florian started coding at 12, developed ML models by 15, and Philip brings ML engineering experience from top companies like Microsoft, McKinsey and BMW. ✨ Antoine Leboyer (Managing Director, TUM Venture Labs AI & Software) A Harvard MBA and seasoned tech executive, Antoine leads an innovation hub at TUM Venture Lab dedicated to supporting tech-based startups in AI and software. His mission? Empower entrepreneurs to translate research into impactful ventures. Meet members from all partner initiatives including CDTM, Manage and More by UnternehmerTUM, START Munich, TEG | The Entrepreneurial Group, Enactus, Innovis VC, MUNICH INNOVATION ECOSYSTEM, Med-Dev, OneAIM, Young Founders Network e.V. and Gründungsmagnet. Spaces are filling fast, so RSVP now to ensure you’re part of the conversation shaping the future of AI in Munich! Let’s connect, learn, and inspire each other! 💡🤖 👉 Link in comments below #tumai #ai #artificialintelligence #machinelearning #AIEntrepreneurship #MunichAI #Networking #Innovation #Entrepreneurship #startups

62

5 Kommentare

I will be attending #ICML Monday-Thursday in Vienna, Austria. If you are around and want to chat about large-scale graph learning or sequence learning and how we use them in Meta ping me. #ICML #ICML2024 #Meta #GraphLearning #SequenceLearning

44

🎉 Excited to share our latest work on extracting better features from diffusion models co-led by Stefan Baumann and Kolja Bauer. ✨ Diffusion models are amazing at learning world representations. Their features power many tasks: • Semantic correspondence • Depth estimation • Semantic segmentation 🤔 But have you ever wondered why we extract diffusion features from noisy images? Doesn’t that destroy valuable information? We show it does - and also requires finding correct hyperparameters for every downstream task. We thought, there had to be a better way. And there is. 🚀 With just 30 minutes of task-agnostic finetuning on a single GPU, we eliminate the need for noisy inputs. ✅ No noise ✅ No timestep tuning ✅ Better features, better performance across many tasks We make code and cleaned 🧹 weights available for SD 1.5 and SD 2.1. Have a look now! ⬇️

75

2 Kommentare

I will be speaking at the AIAI Berlin conference on the topic of Automating Computer Vision Workflows with Foundation Models by AI Accelerator Institute along with Mirza Rahim Baig at the Waldorf Astoria on 11th and 12th. AI Accelerator Institute This is one of the biggest AI summits in Germany and I am excited to be there See you there! #artificialintelligence #businessintelligence #leadership #genai #computervision

60

3 Kommentare

IS THERE AN AI / LLM MANIFESTO? I just started co-teaching a class at the BFH Technik und Informatik focussed on giving our engineers hands-on experience with agile. So I thought, instead of the Agile Manifesto, is there something similiar for AI and LLMs? Probably there is. But this was ChatGPT's take on it, which I don't think is half-bad: 1. Human-Centered Design Over Automation for Its Own Sake Prioritize AI systems that enhance human creativity, decision-making, and interaction over purely automating tasks. LLMs should be built to support and complement human work, not to replace it without meaningful improvement. 2. Transparency and Accountability Over Black-Box Models Ensure AI models are explainable and their decision-making processes transparent, rather than deploying powerful but opaque "black-box" systems. This fosters trust and accountability. 3. Ethical Alignment Over Unchecked Innovation Strive for responsible innovation, ensuring that LLMs and AI are aligned with societal values, ethical principles, and legal frameworks, rather than prioritizing speed or scale at the expense of fairness and safety. 4. Continuous Learning and Improvement Over Fixed Solutions Embrace AI models that are capable of adapting, learning, and evolving based on user feedback and changing environments, rather than rigid systems that do not evolve once deployed.

23

5 Kommentare

Tomorrow is the last day to get substantially reduced early bird tickets for AMLD - Applied Machine Learning Days 2025. Here's my ❤️ pitch why it's the #AI conference you cannot miss: 🎯 The content: Almost 60 tracks and hands-on workshops. The tracks bring you face-to-face with the movers and shakers actively shaping AI's future, while the workshops get you coding and building. This year's quality bar was the highest we've seen in nearly a decade! 🌍 The audience: Our audience brings together decision-makers from across the AI landscape: engineers, scientists, business leaders, practitioners, policy makers, educators, entrepreneurs. This diversity creates unique networking opportunities - whether you're looking for collaborators, talent, investors, or clients, they'll be there. The quality of our audience is why so many participants return year after year. 🤝 The unique mix: AMLD is neither a pure scientific meeting nor trade show - it's a unique mix of both. Having started this at EPFL, one of the leading technical universities, we wanted to create something different: a place where breakthrough research meets practical implementation. Countless projects and collaborations have literally started at AMLD. This mix simply works. 🏔️ The setting: The SwissTech Convention Center is simply extraordinary. And when you step out, you're just minutes away from Lac Léman, with breathtaking views of *the* highest peak of the Alps. EPFL's campus has this unique energy, and being here in February means you can plan to hit the slopes, which should be amazing this time of the year!⛷️ ⚡ The timing: Everyone now realizes that AI is reshaping everything. At AMLD, you can go beyond the social media noise to meet and learn from the people who are actually building and deploying these systems. Whether you're deep in the technical challenges or leading strategic decisions: AMLD will be time well spent. Early bird tickets available until tomorrow at https://lnkd.in/exgqpKah AMLD - Applied Machine Learning Days EPFL AI Center EPFL Alumni EPFL School of Computer and Communication Sciences EPFL School of Life Sciences EPFL School of Engineering

85

Happy to share that our next AI seminar at Technology Innovation Institute will be given by Prof. Lenka Zdeborova from EPFL. Prof. Zdeborova will walk us through phase transitions between positional and semantic learning in the attention mechanism using a mathematically tractable dot-product attention model. 📆 Date & time: 25th June, 2024, 3:00pm (GST) 👉🏻 Register at https://lnkd.in/dwJVZibg 👩‍🏫 Lenka Zdeborová is a Professor of Physics and Computer Science at École Polytechnique Fédérale de Lausanne, where she leads the Statistical Physics of Computation Laboratory. She received a PhD in physics from the University of Paris-Sud and Charles University in Prague in 2008. She spent two years in the Los Alamos National Laboratory as the Director's Postdoctoral Fellow. Between 2010 and 2020, she was a researcher at CNRS, working in the Institute of Theoretical Physics in CEA Saclay, France. In 2014, she was awarded the CNRS bronze medal, in 2016 Philippe Meyer prize in theoretical physics and an ERC Starting Grant, in 2018 the Irène Joliot-Curie prize, in 2021 the Gibbs lectureship of AMS and the Neuron Fund award. She was an editorial board member for the Journal of Physics A, Physical Review E, Physical Review X, SIMODS, Machine Learning: Science and Technology, and Information and Inference. Lenka's expertise is in the application of concepts from statistical physics, such as advanced mean field methods, the replica method, and related message-passing algorithms, to problems in machine learning, signal processing, inference, and optimization. She enjoys erasing the boundaries between theoretical physics, mathematics and computer science. #attention #transformers #LLMs #AI #seminar #theory

70

This week, Dominik Werner Wolf and I presented our work as co-authors at the 46th edition of the German Conference on Pattern Recognition (GCPR 2024) at the Technical University of Munich. It was co-organised by the Munich Center for Machine Learning and the German Association for Pattern Recognition (DAGM). Our work focuses on how various metrics that attempt to estimate the calibration error (present in a model’s predictions) are incoherent in the results they produce and how this raises concerning questions. Since calibration is directly tied to the reliability and trustworthiness of any AI model (be it in the NLP or CV domain), we believe our work is interesting and possesses future scope especially in safety-critical domains like autonomous driving. The work was done as part of my master- thesis research at the Glass Laboratory and Advanced System Solutions (GLASS Lab) at Volkswagen Group’s R&D wing at Wolfsburg. Big thanks to Dominik Werner Wolf for being an amazing mentor throughout my 1.5 years at VW and also for collaborating with me for this publication! I thank Prof. Markus Ulrich and Prof. Alexander Braun too, for their valuable guidance and comments on our work. If you’re further interested in our results, check out our paper on arXiv and provide us your comments/feedback! https://lnkd.in/gFt3Ua2S

68

9 Kommentare

Fraunhofer's OpenGPT-X project releases Teuken-7B on Hugging Face ✅ 7 billion parameters ✅ Trained in all 24 EU languages ✅ 50% non-English data (compared to 8% in LLaMA-3) ✅ Major languages involved: English, French, German, Italian, Spanish. ✅ Context Window: 4K ✅ Open-source & commercially usable ✅ Optimized for multilingual efficiency and cost reduction 🔹 Ideal for European organizations with multilingual needs 🔹 Ensure sensitive data stays secure within your infrastructure 🔹 Customize it for industry-specific applications (e.g., finance, medicine) I can't wait to try it on #Ollama!! Source: https://lnkd.in/d8JbdXc7 Paper: https://lnkd.in/dS5CQZRU Tips: https://lnkd.in/dyQpnSEJ #GenAI #Teuken7B #OpenGPTX #MultilingualAI #Fraunhofer #HuggingFace #OpenSource #EuropeanUnion #EU #TechInnovation #DataSovereignty

14

Exciting new research from TU Wien – Academy for Continuing Education and UCL explores how Large Language Models (LLMs) perform in generating Boolean queries for systematic literature reviews. Here's why this matters: >> Key Findings: - GPT-4 and GPT-3.5 significantly outperformed previous benchmarks in precision scores for the CLEF TAR dataset - Open-source models like Mistral and Mixtral showed competitive performance against proprietary GPT models - The study revealed significant variability in query generation results, highlighting reliability challenges >> Technical Deep Dive: The researchers implemented a comprehensive pipeline that: - Automatically generates Boolean queries from review topics using various LLMs - Tests multiple model variants including GPT-3.5-1106, GPT-4-1106, Mistral-7B, and Mixtral-8X7B - Evaluates queries using PubMed database retrieval - Implements seed-based generation for reproducibility testing >> Model Architecture: The study utilized: - API-based models: GPT-3.5, GPT-4, Mistral-tiny, Mistral-small - Locally-run open source models: Mistral-7B-Instruct-v0.2 and Zephyr-7b-beta - Dense encoder setting using SentenceTransformers for similarity matching This research is crucial for medical researchers and information retrieval specialists looking to automate systematic review processes. The code is publicly available on GitHub for further exploration and validation.

16

Fine-tuning is just one step in the LLM journey. As mentioned in the #DATEV example, fine-tuning was crucial for data anonymization, a necessary pre-phase for analyzing and gaining insights from customer feedback in a way that is compliant with GDPR. Dr. Jonas Rende and Mareike Hoffmann Hoffman's #QED video provides a fascinating look at this process. 📽 https://lnkd.in/dirhgjFx 🤯 Note: While fine-tuning can tailor a model to a specific task, it can also make it less versatile. This specialization can limit its effectiveness in broader contexts. #finetuning #LLM #pragmaticAI #anonymization

8

How to do fun stuff with LLMs: I gave a short introduction to Retrieval Augmented Generation at the University of Basel. Here are some of the slides. Hannes Grassegger had invited me to his media studies seminar. What I wanted to show the students: – How easy it is to talk to ChatGPT and a large text file with a Jupyter notebook and some Python. (Thanks Florian Leuerer for teaching me.) – How this works even easier with a no-code tool like VectorShift, and how to chat with a YouTube video. (Thanks to Nikita Roy.) – How to run a RAG locally, securely and for free on your own computer, with a program like AnythingLLM and different models. (Thanks again, Nikita.) I very much hope that this hands-on training demystified AI to some extent and showed an example of how this partly unreliable technology can be put to good use. Even for journalism. And that you can try out some of these things yourself pretty quickly. (I have already discovered that chunking and sorting results quickly becomes complicated again.)

59

1 Kommentar

📢 New preprint out 🤖🧠🔬   Happy to share our latest work on arXiv: "PEAR: A Robust and Flexible Automation Framework for Ptychography Enabled by Multiple Large Language Model Agents"   🍐 In this study, we introduce PEAR (Ptychographic Experiment and Analysis Robot), a framework that leverages multiple LLM agents to automate ptychographic data analysis workflows. Key features include: - Custom knowledge bases for domain-specific expertise - Human-in-the-loop integration for feedback & control - Multi-agent design for improved accuracy and robustness - Flexible automation levels to suit various user needs   🔎 We demonstrate PEAR's effectiveness through computational experiments and a case study in electron ptychography of 2D SnSe. Our goal is to improve the efficiency and accessibility of ptychography, a powerful computational imaging technique used across many scientific fields. We hope PEAR can contribute to advancing ptychographic analysis and inspire similar approaches in broad computational imaging field and beyond.   🌟 Feedback and discussions are welcome! You can find the full preprint here: https://lnkd.in/gQxpZfAF #Ptychography #MachineLearning #ScientificComputing #AI4Science

16

Whenever you need an inspiration for how to visualize your data, take a look at the work of Matthias Stahl

2

🎉 𝗛𝗮𝗽𝗽𝘆 𝘁𝗼 𝘀𝗵𝗮𝗿𝗲 𝗺𝘆 𝗳𝗶𝗿𝘀𝘁 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝘄𝗼𝗿𝗸! 🎉 Over the past months, I’ve had the pleasure of working with Alexander Heinlein and Eric C. Cyr on a very interesting project that has resulted in our preprint, "DDU-Net: A Domain Decomposition Based CNN for High-Resolution Image Segmentation on Multiple GPUs." This work, which builds upon and expands the results of my master’s thesis, is now available as a preprint on ArXiv! The work addresses the challenging task of high-resolution image segmentation, where using spatial contextual information across different scales is crucial for accuracy. We introduce the DDU-Net (Domain-Decomposition U-Net), a novel CNN architecture, based on the U-Net, where the full high-resolution image is divided into smaller subimages that can be processed largely independently on separate devices, allowing for efficient parallelization. To maintain global context, we incorporated a communication network that facilitates the exchange of low-resolution feature maps between the different devices. 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀: • DDU-Net partitions images into non-overlapping patches for independent processing on multiple GPUs. • A communication network ensures inter-subimage information exchange, preserving spatial context from other subimages. • Our model achieves performance equivalent to a baseline U-Net trained on the full image, but with improved efficiency and scalability. This work wouldn’t have been possible without the insightful discussions with Alexander and Eric. Thank you both for the great collaboration! 📝 Preprint: https://lnkd.in/eTkUbiBT 📚 Thesis: https://lnkd.in/eV-Mu69U #machinelearning #AI #deeplearning #imagesegmentation #highperformancecomputing #parallelcomputing

76

1 Kommentar

✨ Exciting insights from Yann Dubois's recent lecture at Stanford University on building Large Language Models (#LLMs) 🟢 Components of LLMs: Focus on architecture, training algorithms, data collection, evaluation methods, and system optimization for efficient model deployment. 🟢 Pre-training vs. Post-training: Understand the classical language modeling paradigm and the emerging trend of fine-tuning models like ChatGPT to function as AI assistants. 🟢 Importance of Tokenization: Effective tokenization strategies are crucial for handling diverse languages and ensuring model performance. 💎 This lecture is a must-watch for anyone interested in #AI advancements! 👉 https://lnkd.in/dcevp3cJ

44

1 Kommentar