AI Community’s Post

🚀 Fine-Tuning and Hallucinations in LLMs 🤖✨ Here's something counter-intuitive: fine-tuning might make your LLM hallucinate more. Fine-tuning Large Language Models (LLMs) might come with a hidden cost: the introduction of new knowledge during this process could lead to increased hallucinations. 🔍 Study Highlights: 1️⃣ Learning Efficiency: Researchers found that LLMs struggle to acquire new factual knowledge through fine-tuning. Examples introducing new information are learned significantly slower than those consistent with pre-existing knowledge. 2️⃣ Increased Hallucinations: As LLMs eventually learn new facts, their tendency to hallucinate increases linearly with the proportion of new knowledge introduced. 3️⃣ Knowledge Utilization: Fine-tuning primarily teaches LLMs to use their pre-existing knowledge more efficiently, rather than expanding their knowledge base. 💡 - Fine-tuning with new knowledge may encourage models to generate factually incorrect responses, as they are trained to generate facts not grounded in pre-existing knowledge. 💡 - Early stopping and filtering out examples introducing new knowledge can minimize the risk of hallucinations. 📈 Performance Metrics: - Training Dynamics: Unknown fine-tuning examples are fitted substantially slower than known ones, indicating LLMs’ difficulty in integrating new knowledge. - Impact of Knowledge Categories: Fine-tuning on highly known examples yields sub-optimal results, while incorporating examples with lower degrees of certainty (MaybeKnown) enhances performance. 🌟 - Enhanced Stability: Understanding the impact of new knowledge on LLMs helps in designing more stable and reliable models. 🌟 - Efficient Fine-Tuning: Adopting strategies to mitigate hallucinations can lead to more effective use of LLMs in various applications. Stay tuned for more insights on optimizing LLM fine-tuning! 🚀📊 Paper: https://lnkd.in/eHjzq4DF #AI #MachineLearning #LLM #FineTuning #TechNews #Innovation #DataScience

First Research Paper Published! From Knowledge to Action: Leveraging Retrieval Augmented Fine Tuning (RAFT) to Empower Quick and Confident First-Aid Decisions in Emergencies This paper co-authored by me and Kaviru Hapuarachchi explores the potential of Retrieval-Augmented Fine-Tuning (RAFT) to revolutionize first-aid training and application. By integrating RAFT with large language models (LLMs), we propose a novel approach to empower individuals with accurate and situation-specific first-aid instructions during emergencies. RAFT allows AI models to access first-aid knowledge bases, enhancing the accuracy and relevance of first-aid responses in critical situations. The research was part of the Generative AI Challenge by IEEE IES, and I'm incredibly grateful for the opportunity to participate. This research experience has been incredibly valuable. I've learned a tremendous amount about academic writing, conducting research, and the power of collaboration. A huge thank you to our mentors for their invaluable guidance and support throughout this project. I'm excited to continue exploring the potential of AI for positive social impact! Link to paper: https://rb.gy/u2bmz5 Feel free to discuss the paper or ask any questions in the comments below! #FirstAid #AI #RAFT #GenerativeAI #IEEEIES #AI #RAG #LLM

🌟 Reflecting on a Learning Milestone: Unveiling the Art of Research Paper Understanding through Chatting! 🚀🤖 🚀🤖 I am thrilled to share a significant step in my machine-learning journey. In the showcased video, explore the profound insights gained through my interaction with the Gemini-Pro API. This powerful tool seamlessly interacts with research papers in PDF format, enabling me to pose questions and receive articulate responses. This experience has not only deepened my understanding of intricate academic content but has also opened up exciting possibilities for future exploration and growth. 🚀📚 As I reflect on this learning, I am enthusiastic about the future scopes that lie ahead in leveraging such advanced technologies to push the boundaries of knowledge and innovation in the field of machine learning. 💡🌐 Follow my works on GitHub: https://lnkd.in/gjBR7Fzy #MachineLearning #AI #GeminiProAPI #Research #ArtificialIntelligence #LearningJourney #Milestone #Innovation #Technology #CareerDevelopment #KnowledgeAdvancement

🚨 THIS IS INSANE 🚨 New algorithm can help understand animal communication 🤯 It discovers language just by watching videos 🐘 🐳 Scientists at MIT have developed #DenseAV, an algorithm that learns language by watching videos, without needing text input!! Basically, DenseAV processes audio and visual data separately, linking sounds with images through contrastive learning, allowing the model to recognize fine-grained details and perform better than any other algorithm existing 🚀 This could be super useful in multimedia search, language learning, and even understanding animal communication 😳 Check out the paper yourself: https://lnkd.in/dG-SFsiN #ai #artificialintelligence #deeplearning #llms #largelanguagemodels #generativeai #genai #machinelearning #language #mit #data #datascience #aimodels #aidevelopment #aiengineering #audio #video #multimedia #languagelearning

AI and the future of critical thinking It is certainly worth thinking of “intelligence” considering the fact that AI is very smart statistical pattern detection. That means it can only find and generate what was statistically existing before including word relationships. “[…] in a highly recommended essay entitled ‘Now the humanities can disrupt AI,’ Lauren M. E. Goodlad and Samuel Baker write: “As unloved corporate behemoths try to pass off data-scraping statistical models as AI genies, the world’s humanists, composition instructors and creative writers might just be the new MVPs in the struggle for the future of critical thinking.” […]” Here from this article: “A human toddler usually requires just a few examples to recognize that a kangaroo is not an elephant, and that both real-world animals are different than, say, pictures of animals on a sippy cup. And yet, the powerful statistical models now driving “artificial intelligence” (AI)—such as the much-discussed large language model ChatGPT—have no such ability. The human brain evolved over 500 million years to help people make sense of a world of multifarious objects, within the lived contexts that embed their learning in social relations and affective experiences. Deprived of any such biological or social affordances, today’s machine learning models require arsenals of computer power and thousands of examples of each and every object (pictured from many angles against myriad backgrounds) to achieve even modest capabilities to navigate the visual world.” #human #humanities #ai #genai #statistics #patterns #thinking

Power of Deep Focus in Web3 and AI Have you ever felt like you were spreading yourself too thin across multiple topics? In vast fields like Web3 and artificial intelligence, there are countless areas to explore - but true mastery takes focus. This week, let's discuss how laser focusing your learning on one technology, application or skill within these domains can transform your journey. By drilling deep instead of broadening your scope, you develop grit and expertise that lets you ride the waves of change. Diving into one niche topic allows you to truly understand the underlying concepts and develop meaningful hands-on experience through intensive projects. Over time, this accumulates your knowledge to influential levels. Others turn to you as the local expert! Rather than dabbling in a bit of everything, focus your energy into deeply understanding one aspect of Web3 or AI - whether smart contracts, machine vision or natural language processing. Immerse yourself fully and ride that wave all the way to shore. What area could you see yourself exploring with laser focus? Have you noticed your learning accelerate when narrowing your focus? Share your thoughts below! Commit to deep diving into one niche and becoming a true master of your domain.  #Web3 #AI #Mastery #machinelearning

Exciting News: Introducing AI-Powered Apps that Fast-Track Literature Reviews! 🚀 In today's fast-paced research landscape, staying ahead of the curve is crucial. That's why I'm thrilled to share the latest innovation in the world of academia and research – AI-powered tools that are revolutionizing literature reviews. 📚 Efficiency Redefined: These AI-driven tools are designed to streamline the literature review process, making it quicker and more efficient than ever before. Say goodbye to hours of manual searching and sifting through endless documents! 🤖 Smart Recommendations: These apps leverage the power of AI and natural language processing to provide intelligent recommendations, ensuring you discover the most relevant and up-to-date research in your field. 🌐 Global Reach: With the ability to search across a vast range of online databases, these apps break down geographical barriers, giving you access to a global repository of knowledge at your fingertips. 📈 Stay Competitive: In the age of information overload, these AI apps are your secret weapon for staying competitive. Stay on top of emerging trends and breakthroughs in your field with ease. 🧩 Seamless Integration: Many of these apps seamlessly integrate with popular reference management tools, making it easier than ever to organize and cite your sources. Whether you're a researcher, scholar, or student, these AI-powered apps are here to enhance your literature review process. Embrace the future of research and join me in exploring the endless possibilities that AI has to offer. Ready to supercharge your literature reviews? Get started with AI today and unlock a world of knowledge like never before. #AI #Research #LiteratureReview #Innovation #Academia Let's connect and discuss how AI is transforming your research journey! 🤝 #LinkedIn #Networking #AIInnovation

🌍 Excited to share our latest paper on the powerful role of #AI and #KnowledgeGraphs in analyzing #SocialMedia! Explore how AI is revolutionizing our understanding of the complex discussions on micro-blogging platforms. 👉 The paper: https://lnkd.in/dEReCHK3 Recently, many methods and pipelines have emerged for automatically extracting knowledge graphs from documents like scientific publications and patents. However, adapting these methods to include alternative text sources such as micro-blogging posts and news has been challenging, as they struggle with open-domain entities and relations. In this paper, we introduce an advanced information extraction pipeline designed to extract knowledge graphs with open-domain entities from social media posts. Our pipeline uses dependency parsing and classifies entity relations through hierarchical clustering over word embeddings. We demonstrate our approach by extracting semantic triples from a dataset of 100,000 tweets about digital transformation and publicly releasing the resulting knowledge graph. Through two experimental evaluations on the same dataset, we show that our system produces triples with over 95% precision and outperforms similar pipelines while generating a higher number of triples. #DataScience #MachineLearning #SocialMedia #AIApplications #AICommunity #AIAdvancements #DataScience #MachineLearning #DeepLearning #AIprogress #Science #Research #ScienceForAll

𝑾𝒐𝒓𝒍𝒅 𝑴𝒐𝒅𝒆𝒍 𝒐𝒏 𝑴𝒊𝒍𝒍𝒊𝒐𝒏-𝑳𝒆𝒏𝒈𝒕𝒉 𝑽𝒊𝒅𝒆𝒐 𝒂𝒏𝒅 𝑳𝒂𝒏𝒈𝒖𝒂𝒈𝒆 𝒘𝒊𝒕𝒉 𝑹𝒊𝒏𝒈𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 🌐 Exploring the Fusion of Text and Vision in AI: I've just come across an intriguing paper that sets a new precedent in integrating textual knowledge and the physical world. The study, rooted in the advancements of Llama 2, leverages both the Books3 dataset and diverse visual footage for training. This holistic approach aims to understand complex narratives and scenes by jointly training with text and vision; 🔗 Efficient Scaling with RingAttention: The paper introduces a cutting-edge solution to the prohibitive resource demands of long-form text and video understanding. By employing RingAttention, the model efficiently scales the context size from 4K to an impressive 1M. This is further complemented by Rotary Positional Embeddings (RoPE); 📦 Interesting Training Techniques: * Masked sequence packing to accommodate various sequence lengths within a multimodal context; * Loss re-weighting to ensure balanced training across different modes. 🚀 Two-Step Training for Groundbreaking Results: The paper outlines a two-step training process: 1. Extends the context to 1M, employing a power-2 scaling law for gradual expansion. The trained models achieve near-perfect accuracy in needle retrieval tasks up to a 1M context window and are further fine-tuned with a custom-made QA dataset for chat-like behavior; 2. Integrates vision capabilities, mixing text, images, and long-range videos. The resulting models can understand and jointly generate very long sequences of either modality, achieving SOTA in multiple tasks. Really excited to see how this will advance multi-modal large models further! 🤗 Project page: https://lnkd.in/dYJJxzmF Paper: https://lnkd.in/duUA2aUT Models: https://lnkd.in/dCkm46Xn #AIResearch #Technology #AI #LanguageModels #ArtificialIntelligence #Innovation #DataScience

📢Excited to share the second tutorial on "How to Ask Queries from AnswerThis.io?" We will be diving into the fundamentals of asking research queries, conducting a literature review, and finding answers to any doubts you may have with AnswerThis.io. Let us get started! ✔Sign in and ask your queries, followed by the adjustment of filters. ✔Receive an answer that is in-line cited and supported by pertinent literature in the form of evidence-based sources that you can easily add to your reference manager. ✔By selecting the sources that resonate with you, you can get concise summaries to enhance your understanding. 📝We greatly appreciate your feedback in the comments! ✨Coming soon: the next tutorial! #ai #artificialintelligence #aiinresearch #research #literaturereview #scientificwriting #publication #Answerthis #AIResearchAssistant