Gianvito Grasso

#AI Breakthroughs: #Models with #Tiny or #small #Datasets in #medical #image #analysis I’m delighted to present our latest publication on "Hard Exudates Segmentation in Diabetic Retinopathy, utilizing the DiaRetDB1 dataset". Our study showcases how we achieved outstanding accuracy with a compact dataset of just 89 images. This breakthrough research, conducted in collaboration with the exceptional @Ma Yinghua(China), introduces new techniques for advancing medical image analysis. The full paper is available for you to explore here: #IEEE #Xplore Document.

14

Streamline your research workflow with our AI-powered Automated Quality Control. Seamlessly embedded in Concentriq for Research, the module: ▪ Identifies artifacts 6x faster than manual review. ▪ Detects 10 common artifact types - including tissue folds, air bubbles, and out of focus areas - with configurable sensitivity settings. ▪ Provides results that can be exported to image analysis applications. Want to learn more? Contact us here: https://bit.ly/3QIRupq #digitalpathology

17

🚀 Cut Inference Costs with UbiOps 💬 Significantly improve the accuracy of your #GenAI models whilst reducing inference costs through fine-tuning and utilising #RAG. 🔧 Fine-Tuning: Customise models with industry-specific data, enhancing accuracy. 🔍 RAG: Helps add relevant context by employing vector databases, reducing re-prompting and model hallucinations. Remember, accurate models mean fewer prompts and reduced costs! Discover the full guide here: https://bit.ly/3Ywbeky #AI #MachineLearning #FineTuning #RAG #InferenceCosts #UbiOps #RAG #MLOps

4

Fabian Theis and team released yesterday a masterpiece benchmark using single cell data in a Nature Machine Learning paper about self-supervised models using a vast amount of non-labelled data, in this insane case with more than 20M cells. Their outcomes for brain cell types were surprisingly as good as traditional supervised methods ~82-83% against ~36% in baseline model using F1 score. To run this benchmark his team used a super GPU cluster config (multiple V100 and A100 cards), reaching 160GB in 2-days use. What is a little bit expensive to run in the cloud 🫠. Congrats Fabian and team! Amazing work. Wanna more details? Check it out in this link: https://lnkd.in/dK4mdWUH

72

1 Kommentar

We are thrilled to announce that two papers from our team and collaborators have been accepted to the NeurIPS 2024 Datasets & Benchmarks Track! 🏆 1️⃣ HARMONIC: Harnessing LLMs for Tabular Data Synthesis and Privacy Protection   This work introduces a new framework for generating privacy-preserving synthetic tabular data using LLMs. HARMONIC’s novel approach enhances privacy while maintaining data utility, using instruction fine-tuning and k-nearest neighbors to train models to understand data structure without risking privacy leakage.   Co-authors: Yuxin Wang, Duanyu Feng, Yongfu Dai, Zhengyu Chen, Jimin Huang, Sophia Ananiadou, Qianqian Xie, and Hao Wang. 📄 https://lnkd.in/ezNETwdX 2️⃣ FinBen: A Holistic Financial Benchmark for Large Language Models   FinBen is the first comprehensive open-source financial benchmark, featuring 36 datasets across 24 tasks, from stock trading to risk management. It evaluates 15 top LLMs, including GPT-4, and provides a robust platform for advancing financial LLM research.   Co-authors: Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, Duanyu Feng, Yijing Xu, Haoqiang Kang, Ziyan Kuang, Chenhan Yuan, Kailai Yang, Zheheng Luo, Tianlin Zhang, Zhiwei Liu, Guojun(国钧) Xiong, Zhiyang Deng, Carolyn(Yuechen) Jiang, Zhiyuan Yao, HaoHang Li, Yangyang（Shirley） Yu, Gang Hu, Huang Jiajia, Xiao-Yang Liu, Alejandro Lopez-Lira, Benyou Wang, Yanzhao Lai, Hao Wang, Min Peng, Sophia Ananiadou, and Jimin Huang. 📄 https://lnkd.in/ehMvQyap and the leaderboard: https://lnkd.in/ewyAgbUd These papers highlight the incredible potential of LLMs in privacy protection and financial decision-making, and we’re proud to be part of such innovative work. A huge congratulations to all co-authors involved! 🎉 #NeurIPS2024 #LLMs #Datasets #FinBen #HARMONIC #TheFinAI #AIResearch #PrivacyProtection #FinancialNLP #SyntheticData

6

Mathematic legend Terence Tao is playing with OpenAI latest model 🍓: 📌 “[it is] more capable tool than previous iterations, though still struggling with the most advanced research mathematical tasks” 📌 “the new model could work its way to a correct (and well-written) solution *if* provided a lot of hints and prodding, but did not generate the key conceptual ideas on its own, and did make some non-trivial mistakes.” 📌 “The experience seemed roughly on par with trying to advise a mediocre, but not completely incompetent, graduate student.” 🤣 📌 Here is the interesting one: “It may only take one or two further iterations of improved capability […] until the level of "competent graduate student" is reached, at which point I could see this tool being of significant use in research level tasks” 📌 “I could imagine a model of this capability that was specifically finetuned on Lean and Mathlib, and integrated into an IDE, being extremely useful in formalization projects” Even if Terence Tao is not saying “it’s revolutionary” or “I used 🍓 and solved the riemann hypothesis” … I think his review is quite posititve. Remember, we started in Dec 2022 with a “kindergarten model” #GPT3 … who became a mediocre grad student in less than 2 years. To me it sounds almost like “Sheldon Cooper” precocity (#BigBangTheory) 🤩😎 PS1 - Marcel Salathé : thank you for sharing the source material! PS2- Picture below is not of Terence Tao but of “a mediocre grad student holding a strawberry” accroding to ChatGPT 🤣 Link in the comments section 👇

14

1 Kommentar

𝐁𝐢𝐨𝐌𝐢𝐬𝐭𝐫𝐚𝐥 - 𝐎𝐩𝐞𝐧 𝐋𝐋𝐌 𝐟𝐨𝐫 𝐁𝐢𝐨𝐦𝐞𝐝𝐢𝐜𝐚𝐥 𝐃𝐨𝐦𝐚𝐢𝐧 BioMistral is an open-source LLMtailored for the biomedical domain BioMistral is based on Mistral LLM and further pre-trained on PubMed Central text corpus. BioMistral is evaluated on a benchmark comprising 10 established medical question-answering (QA) tasks in English. BioMistral outperforms existing LLMs like PMC-Llama, MedAlpaca, MediTron etc. BioMistral LLM is available on Hugging Face platform (details in the comments) #llms #biomedical #generativeai #ai #nlproc #deeplearning

112

1 Kommentar

🚀 Introducing Neural Scribe's Dictation Feature – AI-Powered Workflow Efficiency! 🚀 🗣️ What is Dictation? Neural Scribe’s Dictation converts spoken recordings into precise clinical notes, including HPI, drastically reducing manual transcription time. 📢 Quick Phrase Tool: Set custom phrases triggered by keywords during dictation (e.g., "insert diet"). Instantly insert preset phrases, like "Take care, this is Dr. James," with a simple command, saving time and reducing repetitive tasks. 🤔 Why Neural Scribe? - AI-powered dictation for accurate documentation - Quick Phrase tool for increased efficiency - Automatic clinical notes and medical coding - Saves 60-80 hours per month for patient care 📽️ Watch this video here: https://lnkd.in/eSK3P49j ⭐ Let’s Transform Your Practice - Contact us or comment below to see how we can help automate your non-billable tasks, freeing up 60-80 hours per month to focus on patient care. #Healthcare #HealthInnovations #HealthcareIT #AI #MedicalScribe

5

Med-Gemini Google released Med-Gemini, a remarkable advancement in AI technology tailored specifically for medical applications. What makes Med-Gemini stand out? 1. Mastering Multimodality: Tackling complex medical data effortlessly, thanks to its advanced reasoning and understanding of multimodal inputs. 2. Setting New Standards: Surpassing benchmarks and even outperforming GPT-4 models across various medical tasks, including diagnostic challenges and medical question answering. 3. Real-World Impact: Demonstrating superiority over human experts in tasks like medical text summarization and referral letter generation, with promising applications in medical research, education, and dialogue. The Med-Gemini models were enhanced from the Gemini family by focusing on three main areas: 1. Advanced Reasoning via Self-Training and Web Search Integration: Two versions were developed, Med-Gemini-M 1.0 and Med-Gemini-L 1.0, tailored for different levels of language tasks. These models were trained to utilize web searches and incorporate a new intelligent search method for solving complex medical problems. 2. Multimodal Understanding via Fine-Tuning and Customized Encoders: Med-Gemini-M 1.5 was fine-tuned to better comprehend various types of medical data (multimodal), while Med-Gemini-S 1.0 was introduced to handle novel types of medical information using specialized tools. 3. Long-context processing with chain-of-reasoning: Med-Gemini-M 1.5 was adjusted to handle lengthy medical records more effectively. Additionally, a new approach was developed to assist these models in better understanding these extended narratives. Paper: https://lnkd.in/gbMnYPwn #LLM #GenAI #MedGemini #Gemini #LanguageModels #AIinMedicine #GPT4 #Medicine

11

3 Kommentare

I just read an interesting paper on 𝘔𝘦𝘨𝘢𝘈𝘨𝘦𝘯𝘵: 𝘈 𝘗𝘳𝘢𝘤𝘵𝘪𝘤𝘢𝘭 𝘍𝘳𝘢𝘮𝘦𝘸𝘰𝘳𝘬 𝘧𝘰𝘳 𝘈𝘶𝘵𝘰𝘯𝘰𝘮𝘰𝘶𝘴 𝘊𝘰𝘰𝘱𝘦𝘳𝘢𝘵𝘪𝘰𝘯 𝘪𝘯 𝘓𝘢𝘳𝘨𝘦-𝘚𝘤𝘢𝘭𝘦 𝘓𝘓𝘔 𝘈𝘨𝘦𝘯𝘵 𝘚𝘺𝘴𝘵𝘦𝘮𝘴. It introduces MegaAgent, a framework designed to improve how Large Language Model (LLM) multi-agent systems work, especially when it comes to autonomy and scalability. 𝗧𝗵𝗲 𝗰𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲: Many current LLM-based systems rely heavily on predefined procedures, which can be limiting when dealing with complex, real-world tasks. These systems often struggle to manage large numbers of agents and ensure they work well together. 𝗠𝗲𝗴𝗮𝗔𝗴𝗲𝗻𝘁’𝘀 𝗮𝗽𝗽𝗿𝗼𝗮𝗰𝗵: This framework takes a different approach by dynamically creating agents based on the needs of the task at hand. It divides tasks automatically and manages multiple operations at once, all within a structured system designed to scale effectively. 𝗪𝗵𝘆 𝗶𝘁’𝘀 𝗻𝗼𝘁𝗲𝘄𝗼𝗿𝘁𝗵𝘆: The effectiveness of MegaAgent was demonstrated in two experiments—one involving a game development task and another simulating national policy-making. In both cases, MegaAgent managed to handle up to 590 agents and completed tasks more efficiently than other systems. 𝗛𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲 𝗜𝗺𝗽𝗮𝗰𝘁: While the primary focus of MegaAgent may not be healthcare, its capabilities in managing large-scale, complex tasks and facilitating effective collaboration could be applied to improve various aspects of healthcare delivery and policy-making. Personalized medicine, healthcare policy simulation, multi-disciplinary collaboration, operational efficiency, and patient monitoring could greatly benefit from it. 𝗧𝗵𝗲 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆: MegaAgent could offer new ways to manage large-scale, multi-agent tasks more effectively, especially for those working with LLMs. I am curious to see how this framework might influence future developments in AI and healthcare. https://lnkd.in/gmgpvTYE #AI #LLM #MultiAgentSystems #Research

3

1 Kommentar

Can neural networks be turing complete? Traditionally, without CoT (chain of thought), transformers are limited to parallelizable tasks within the AC0 complexity class, solving problems computable by shallow Boolean circuits due to their fixed architecture and inherent parallelism. CoT reasoning fundamentally changes this landscape by enabling transformers to handle sequential computations through intermediate reasoning tokens. Each CoT step effectively adds a layer to the computational graph, increasing computational depth and allowing the model to simulate deeper circuits beyond AC0. This advancement brings transformers into the realm of P/poly—the class of problems solvable by polynomial-size circuits. With sufficient CoT steps, a transformer could, in theory, simulate any computation that a polynomial-size circuit can perform, narrowing the gap between transformers and Turing machines. Practical limitations exist, such as finite context windows and computational resources. Leveraging this potential requires careful model design and optimization. Extending this concept, we can envision neural networks functioning analogously to operating systems. Just as an OS manages resources and orchestrates complex processes, neural networks could dynamically control computational tasks, allocate attention, and sequence operations to solve intricate problems efficiently. Integrating CoT reasoning in transformers offers a promising path toward Turing completeness. Despite practical challenges, this perspective could revolutionize artificial intelligence by transforming neural networks into general-purpose computational frameworks capable of complex reasoning and decision-making. An excellent paper

12

The FIRST NEURAL NETWORK It is well know that neural networks are the modular unit of #AI, and are powering the ongoing technological revolution. However, what many people do not realize is that the very first neural network dates as far back a 1805. In 1805, Adrien Marie Legendre published a method which given a set of data points, determines the line of best fit in the least squares sense. The so called "method of least squares" minimized the sum of squared distances between the data points ("training data") and the candidate lines. Carl Friedrich Gauss published the same idea in 1809 but noted that he had the result since as early as 1795. The method of least squares, commonly known today as "LINEAR REGRESSION" is indeed the first known fully end-to-end differentiable method that learns a model from training data. And since that is exactly the defining property of neural networks, it follows that linear regression is the first known neural network. #artificialintelligence #mathematics

223

17 Kommentare

Exploring LaRS @ SWISSUbase : https://lnkd.in/e6YKynrx. The Language Repository of Switzerland LaRS offers a specific file and #metadata structure for the studies deposited on the platform. In the sixth post of the ShareTIGR lab blog, we look into the details of this structure and into their implications for the searchability and findability of the #data records stored on the repository. #dh #fairdata #openresearchdata

19