Efstratios Gavves’ Post

🌐Why Transformers Are Game-Changers: The Secrets Behind 'Attention Is All You Need Ever wondered how models like BERT and GPT became so good at understanding language? The answer lies in self-attention—and it’s simpler than you might think! I’ve just written a beginner-friendly guide on Transformers and the magic behind “Attention Is All You Need.” Whether you're a tech enthusiast or just curious about how these AI models understand and generate text, this blog is for you! 🔧 What you’ll get from my latest blog: A super easy explanation of self-attention (with real-life examples!) How Transformers changed the game in AI and why they’re better than older models A breakdown of the Encoder and Decoder architecture without the heavy jargon! If you've ever been confused by the term “self-attention” or want to see how Transformers really work, check out the guide here 👉 https://lnkd.in/gWB3Ekyf

Just finished the course “RAG and Fine-Tuning Explained” by Morten Rand-Hendriksen! An excellent short course and a brilliant idea to present it as a 20-minute session, addressing the challenge of defining RAG and clarifying the capabilities and limitations of LLMs. Check it out: https://lnkd.in/eZggXdZ5 #generativeai #aisoftwaredevelopment #retrievalaugmentedgeneration.

https://lnkd.in/djES92BA New, powerful open-source large language models (LLMs) are shaking things up! These models, like Gemma, Phi, Zephyr, Mistral, and Llama, are more affordable and give organizations control over the underlying tech. But, they might not be quite as good as some commercial options. That's where fine-tuning comes to the rescue. By fine-tuning these open-source models, you can get performance close to that of GPT-3, all while keeping costs down and owning the model itself.

Just finished the course “Advanced AI: Transformers for Computer Vision” by Jonathan Fernandes! Check it out: https://lnkd.in/gJ4BpKZh #generativeai #transformermodels #generativeadversarialnetworks.

Models that function as glue between specific models will allow us to achieve performance across less popular modalities. Once that happens in large, the capabilities of AI will improve dramatically. You can read my articles for free by following the IAEE LinkedIn. Machine Learning - IAEE

Phi-4 shows what’s next in AI. Despite its compact size, Phi-4 outperforms significantly larger models such as GPT-4o (its teacher model!) and Gemini Pro 1.5 in areas like mathematical reasoning. What’s fascinating about Phi-4 is its remarkable training strategy. Even after several epochs on synthetic data, the team observed no overfitting. In fact, the 12-epoch version surpassed models trained on more unique web tokens. It’s not about the size of the model but the intelligence behind its architecture and training methods!

Just finished the course “Advanced AI: Transformers for Computer Vision” by Jonathan Fernandes! Check it out: https://lnkd.in/eU6QqPNH #generativeai #transformermodels #generativeadversarialnetworks.

Super excited to share that our paper "Towards Calibrated Robust Fine-Tuning of Vision-Language Models" has just been accepted at #NeurIPS2024 🎉🎉🎉 📑 arXiv: https://lnkd.in/gqsb5bhh 💻 code: https://lnkd.in/gSB7pJd6 💥 highlight: We found that existing fine-tuning approaches hurt the uncertainty calibration of the foundation model, e.g., CLIP, especially on out-of-distribution (OOD) data! To ensure uncertainty calibration as well as robustness on OOD, we derived a new theoretical bound over OOD calibration and classification errors, which is empirically estimated with only the in-distribution data!! and propose a practical implementation to optimize that bound for fine-tuning of VLMs. 🙌 It was my honor to collaborate with exceptional scholars -- Hyesu Lim (co-first author), Mijoo Kim, Dongyoon Han, Sangdoo Yun, Jaegul Choo, Alex Hauptmann, Zhi-Qi Cheng, and Kyungwoo Song -- from diverse institutions (Yonsei University, Carnegie Mellon University, KAIST AI, Chung-Ang University, NAVER Corp) See you in Vancouver 🇨🇦 :)

👋 Hi everyone, For Data Jam episode 13, our topic will be "HCI meets AI/ML paper study: Generative Agents Simulacra". This paper presents generative agents, which are software entities designed to mimic realistic human actions. The framework enhances a sophisticated language model to encode and reproduce agents’ memories, experiences, and intentions using natural language. References: - https://lnkd.in/gWdwsWG7 - https://lnkd.in/dkJsjzMx What is the event format❓️ Basically, we're going to be exploring and studying a specific topic and hopefully it will create a discussion around that topic. See you on our discord server. If you haven't join the server, check the invitation link on below. Discord link: https://lnkd.in/gca6s8ug For Data Jam event reminder, you can register through this link: https://lnkd.in/gKbbGnfx And we will invite you to our google calendar so you won't miss it! #datascience #machinelearning #artificialintelligence

CS50 Artificial intelligence lecture https://lnkd.in/dcd-eU5A Summary The minimax algorithm evaluates moves in games like Tic-Tac-Toe, helping players make optimal decisions based on potential outcomes. Highlights 🤖 Minimax Algorithm: A systematic way to evaluate moves in games like Tic-Tac-Toe. 🧩 Possible Moves: O can make three moves, each leading to different outcomes. 📊 Scoring Outcomes: Assigns scores of 1 (win), -1 (loss), or 0 (tie) based on game results. 🎮 Recursive Evaluation: Considers all potential responses from opponents to determine best moves. 💡 Reinforcement Learning: Computers learn from feedback, enhancing decision-making. 🧠 Neural Networks: Explores deep learning applications in AI, like language models. 🤹♂️ AI Challenges: Discusses the potential and limitations of AI across various fields. Key Insights ⚙️ Minimax Algorithm’s Importance: This algorithm is vital for strategic decision-making in competitive environments, ensuring players can anticipate opponents’ moves. 🌀 Recursive Nature: The recursive evaluation process allows for comprehensive analysis of the game state, leading to informed choices based on future possibilities. 🎯 Outcome Scoring: The scoring system simplifies complex decisions, providing a clear metric for evaluating the effectiveness of different moves. 🧠 Reinforcement Learning Dynamics: This concept illustrates how AI improves over time, adapting its strategies based on success or failure, similar to human learning. 🔗 Neural Networks in AI: Deep learning facilitates advanced AI applications, enhancing capabilities in natural language processing and other domains. 🌍 AI’s Broad Applications: From gaming to meteorology, AI’s potential spans various industries, offering innovative solutions and insights. 😂 Humor in AI Limitations: The light-hearted poem about a Homework Machine underscores that while AI is powerful, it still has its quirks and limitations, reminding us of its human-like imperfections. #ai #CS50 #HamidRaza #python