Presenting TILR paper at IJCAI 2024

🚀 Exciting Update! 🚀 I'm thrilled to share that our latest research paper, "Navigating Challenges and Technical Debt in Large Language Models Deployment," has been accepted for publication in the prestigious EuroMLSys conference and ACM Journal! The paper is a collaboration between the Mastercard AI Engineering Team led by Bijay Kumar and Dr. Pasquale Minervini, PhD from the School of Informatics at the University of Edinburgh. It delves into the complexities surrounding the deployment of Large Language Models (LLMs), shedding light on unique challenges including memory management, parallelism strategies, model compression, and attention optimization. These challenges underscore the necessity for tailored solutions and sophisticated engineering approaches to ensure the seamless integration of LLMs into production environments.   👇The paper is available to read in ACM Journal Library 📚 https://lnkd.in/epGWvzEt   📢 This marks my second post on this topic! I'm excited to share that you can now explore more about our research through the following resources:   Conference Page: https://lnkd.in/d7p5-Nu9 Video Presentation: https://lnkd.in/dZaDGV2F Here's to surmounting challenges, forging new pathways, and shaping the future of LLM deployment! #mastercard #llms #LargeLanguageModels #EuroMLSys #llmops #generativeai #ai #technology

Excited to share that my paper, titled "Boosting Few-Shot Detection with Large Language Models and Layout-to-Image Synthesis", has been accepted to this year’s Asian Conference on Computer Vision (ACCV 2024)! The paper tackles the persistent challenge of limited data availability in few-shot object detection. Highlights: 1. A novel collaborative framework that employs Large Language Models (LLMs) to extrapolate from spatial layouts of objects, and Layout-to-Image Synthesis (LIS) models for high-quality image generation. 2. The introduction of a Layout-Aware CLIP Score (LACS), a scoring mechanism that ensures tight alignment between generated images and their layouts, eliminating noisy samples by detecting out-of-layout hallucinations (shown in picture). 3. Achieving significant improvements on COCO few-shot benchmarks, boosting detection performance by more than 140% in 5-shot scenarios on a YOLOX-S baseline. Generative augmentation holds significant promise for enhancing detection systems, particularly in data-scarce environments. This work highlights the power of combining LLMs with advanced diffusion models to drive practical advancements in computer vision. Special thanks to my amazing co-authors, Nikolas Ebert and Oliver Wasenmüller, for their incredible collaboration and contributions! Paper: https://lnkd.in/gh2bxHZT #ACCV2024 #ComputerVision #FewShotLearning #GenerativeAI #ObjectDetection #AIResearch

🎆 New Special Issue open for submission: #AI in #Game_Theory: Theory and Applications 🎓 Editor: Dr. Alessia Donato and Prof. Dr. David Carfì 🧭 Deadline for manuscript submissions: 30 June 2025 🔗 Link: https://buff.ly/4ewoCK5 🍀 In this Special Issue, we encourage submissions providing possible applications of AI in Game Theory. The topics of interest for this publication include, but are not limited to, the following: ✔️ Artificial Intelligence in Game Theory. ✔️ Artificial Intelligence in decision-making. ✔️ Evolutionary algorithms in Game Theory. #MDPIOpenAccess #ComSciMath_Mdpi

Hoolaa connections!!!!! I have successfully completed the project titled "Performance Evaluation on Real-time Object Detection Using DL Techniques". I gained invaluable experience working with state-of-the-art deep learning models, specifically YOLOv8 and Faster R-CNN. Implementing YOLOv8, known for its efficiency in real-time object detection, allowed me to explore its robustness and accuracy in various scenarios. Concurrently, employing Faster R-CNN provided insights into its effectiveness in precise object localization and classification tasks.. These findings were subsequently documented and published in the esteemed IJCSE journal, ensuring that our research and insights reach a wide audience in the field of computer science and artificial intelligence. #Deeplearning #Detection #YOLOV8 #Fasterrcnn #ResearchPaper

This is one of the clearest analyses of the problems associated with reaching any form of intelligence with LLMs by John Ball. It clearly and convincingly demonstrates that the problem is not one of engineering the current science but one of needing new science to solve many of the fundamental problems that we are all finding with GenAI. "Kurzweil’s definition was clarified as: “AI that can perform any cognitive task an educated human can.” So that certainly requires the machine to solve today’s limitations with LLMs like hallucinations, and the inclusion of lossless knowledge in context!" The rest of the article follows up on this and identifies some of the many scientific problems that need to be solved. He quotes "A few months ago, Thomas Dietterich presented on LLM problems and what is needed, instead." suggesting that we watch how presentation at https://lnkd.in/eddNT4nk (I haven't watched it yet, but the couple of screen grabs are very informative. what is clear from this is that 2029 is definitely not remotely feasible, far too much science has to be started. https://lnkd.in/e2UR7gqE

I recommend this insightful paper on Structured State Space Models (SSMs) to my students and community. SSMs are gaining attention as alternatives to Transformers, particularly for their efficiency in handling long sequences. Key points include: - SSMs and Transformers are related through structured matrices. - The paper introduces Mamba-2, an improved SSM architecture that is faster and competitive with Transformers. - An efficient algorithm is presented, combining the benefits of SSMs and attention mechanisms. For those interested in the latest advancements in AI and language modeling, this paper is worth reading. https://lnkd.in/djbAWq7U #AI #SSM #Transformers #LanguageModeling

Qwen2-Math, a groundbreaking large language model specifically designed to tackle complex mathematical problems. This model has achieved remarkable performance, scoring 84% on the MATH Benchmark, which includes 12,500 challenging competition-level math problems. Qwen2-Math's exceptional capabilities are poised to revolutionize fields such as scientific research, engineering, and finance by automating intricate calculations and enhancing problem-solving efficiency. This advancement underscores the rapid progress in AI, making sophisticated mathematical reasoning more accessible across various industries. #Qwen2 #MachineLearning #ArtificialIntelligence https://lnkd.in/dCnbq4rW

We're ecited to share our latest research work presented at CVPR 2024 Workshop! 🚀 Toghether with my co-authors Masud Fahim and Prof. Jani Boutellier, we tackled the challenge of feature distributoin shifts in deployed video recognition systems, caused by factors like environmental conditions (rain, lighting variations) and data compression. Our solutions is a simple test-time self distillation strategy that allows deep neural network to adapt their predictions on the fly. This method also works with various modelling architecture (CNNs, transformers, or hybrid) and doesn't need access to the training data. We evaluated our system on several benchmark video datasets, i.e. Kinetics-400, SSV2., and showed significant improvements in performance. 🔥 📑 Paper: https://lnkd.in/gmwZvU65 👩💻 Code: Comming soon. #CVPR2024 #MachineLearning #DeepLearning #VideoRecognition #AI #Research

🚀 Thrilled to Share My First Publication! 🚀 I'm excited to announce that my first research paper, titled "Deep Learning for Pothole Detection: Exploring YOLO V8 Algorithm's Performance in Pavement Detection," has been presented at a conference and published by IEEE! 🌐📖 A heartfelt thanks to my guides, Vishnu Sir and Gnana King Sir, for their invaluable support and mentorship throughout this journey. Their guidance was instrumental in achieving this milestone. Looking forward to more research contributions in the future! #Research #DeepLearning #ComputerVision #MachineLearning #AI