Pharmacoevidence’s Post

I am happy to share that I had a profound opportunity to present the "Multi-Stage Prompting for Next Best Agent Recommendations in Adaptive Workflows" paper at the prestigious 38th Association for the Advancement of Artificial Intelligence (AAAI) conference in Vancouver, Canada. A big thank you to Prerna Agarwal and Renuka Sindhgatta for providing me with the opportunity to work on this research project. It was an amazing experience, starting from scratch with the problem statement to delivering the conference presentation, which is just the beginning of many more to come. #AAAI #IBM #ibmresearch #NLP #Research #AI

Exploring the GenAI Frontier: A Journey through Transformers, GPT, and Accelerated Innovation Hello Connections, I'm thrilled to announce the completion of a comprehensive document titled "Navigating the GenAI Frontier: Transformers, GPT, and the Path to Accelerated Innovation"! In this presentation, we delve deep into the foundations and advancements in the field of Artificial Intelligence (AI), uncovering seminal papers that have paved the way for unprecedented innovation. As we embark on this journey through the GenAI frontier, it's imperative to understand the pivotal works that have shaped the landscape of AI. From the groundbreaking advancements in Transformers to the transformative capabilities of Generative Pre-trained Transformers (GPT), each milestone represents a significant leap forward in our quest for intelligent machines. Understanding these foundational works not only enriches our knowledge but also empowers us to drive future advancements and harness the full potential of AI technologies. By standing on the shoulders of giants and leveraging the insights gleaned from these seminal papers, we pave the way for accelerated innovation and groundbreaking discoveries in AI. I would like to express my heartfelt gratitude to Kanav Bansal Sir and Innomatics Research Labs for their unwavering support, invaluable resources, and boundless knowledge. It is through their guidance and mentorship that we are able to navigate the complexities of the GenAI frontier with confidence and enthusiasm. Let's continue to push the boundaries of possibility, embrace the transformative power of AI, and chart a course towards a future where intelligent machines reshape the world as we know it. Innomatics Research Labs, thank you for being the catalyst for our journey into the realms of Generative AI, Machine Learning, Data Science, and NLP. #GenerativeAI #MachineLearning #DataScience #NLP

Excited to share this! Although a bit delayed, I’m excited to post about my participation in GUVI Geek Networks, IITM Research Park's webinar on "Mastering Machine Learning, Computer Vision, and NLP with Gemini AI", which took place in July 2024. It was a fantastic learning experience, where I deepened my understanding of these cutting-edge technologies. Looking forward to applying these skills in real-world projects and continuing to learn more in the AI space! #Webinar #MachineLearning #ComputerVision #NLP #AI #ContinuousLearning #MachineLearning #LifelongLearning

xLSTM - Extended Long Short Term Memory 🤐 #LSTM which got introduced in late 90s is been a very crucial while considering multiple deep learning use cases especially in areas of NLP. But invention of Transformer architecture and self-attention mechanism slowly outdated LSTM. But "𝘏𝘰𝘸 𝘧𝘢𝘳 𝘤𝘢𝘯 𝘸𝘦 𝘨𝘦𝘵 𝘪𝘯 𝘭𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘮𝘰𝘥𝘦𝘭𝘪𝘯𝘨 𝘪𝘧 𝘸𝘦 𝘴𝘤𝘢𝘭𝘦 𝘓𝘚𝘛𝘔𝘴 𝘵𝘰 𝘣𝘪𝘭𝘭𝘪𝘰𝘯𝘴 𝘰𝘧 𝘱𝘢𝘳𝘢𝘮𝘦𝘵𝘦𝘳𝘴 𝘢𝘯𝘥 𝘶𝘴𝘦 𝘵𝘩𝘦 𝘭𝘢𝘵𝘦𝘴𝘵 𝘵𝘦𝘤𝘩𝘯𝘪𝘲𝘶𝘦𝘴 𝘰𝘧 𝘮𝘰𝘥𝘦𝘳𝘯 𝘓𝘓𝘔𝘴, 𝘣𝘶𝘵 𝘮𝘪𝘵𝘪𝘨𝘢𝘵𝘦 𝘵𝘩𝘦 𝘬𝘯𝘰𝘸𝘯 𝘭𝘪𝘮𝘪𝘵𝘢𝘵𝘪𝘰𝘯𝘴 𝘰𝘧 𝘓𝘚𝘛𝘔𝘴?" 🤔 Exponential gating and modified memory structures boost xLSTM capabilities to perform favorably when compared to state-of-the-art Transformers and State Space Models, both in performance and scaling. It introduced two major inventions: ⭐ Firstly, exponential gating with appropriate normalization and stabilization techniques. ⭐Secondly, modified LSTM memory structure: (i) sLSTM with a scalar memory, a scalar update, and new memory mixing, (ii) mLSTM that is fully parallelizable with a matrix memory and a covariance update rule. Both sLSTM and mLSTM enhance the LSTM through exponential gating. Its showing promising results while comparing with transformer models like Llama or Mamba. But yet to see how its perform on real time use cases ... Paper: https://lnkd.in/gHmFzEP8 #lstm #generativeai #datascience #xlstm #llm #nlp

Excited to share that I have completed a comprehensive course on multimodal Retrieval-Augmented Generation (RAG)! Throughout this course, I delved deep into the world of Generative AI. This has equipped me with the skills to develop AI-driven solutions that leverage contextual retrieval and multimodal generation, including both text and images. Key highlights of my learning journey include: Understanding the integration of multimodal retrieval mechanisms with generative models to enhance AI outputs. Exploring various applications of multimodal RAG in real-world scenarios. Gaining hands-on experience in model training and performance optimization. I am thrilled to apply this cutting-edge technology to create data-driven insights and AI-powered applications that drive innovation and efficiency. Looking forward to connecting with fellow AI enthusiasts and professionals to share knowledge and collaborate on exciting projects! #AI #MachineLearning #NLP #GenerativeAI #MultimodalRAG #DataScience #Innovation #AWS

🔄 Unlocking the Power of Gradient Accumulation for LLMs 🔄 Gradient accumulation is a game-changer in fine-tuning large language models (LLMs), offering a smarter way to handle memory and boost training efficiency. Here’s why it’s worth knowing about: Why Gradient Accumulation? 🌟 Memory Efficiency: Train larger models or use bigger batch sizes without maxing out your GPU memory. 🔍 Stable Training: Achieve more stable and accurate model updates by accumulating gradients over several smaller batches before performing a single update. 🚀 Enhanced Performance: Improve model performance and convergence with effective gradient accumulation techniques. Key Hyperparameters: - Batch Total Size: The overall size of the data processed in each epoch, influencing model performance and training speed. - Batch Per Device: How the total batch is divided among available devices, optimizing parallel processing. - Gradient Accumulation Steps: The number of smaller batches to accumulate gradients before updating model weights, crucial for managing memory and stability. - Total Steps: The total number of training steps, determining how long the model is trained and how often weights are updated. Gradient accumulation allows for efficient memory usage and better training dynamics by adjusting these hyperparameters. For more insights into this technique and its benefits, check out our latest work! #MachineLearning #AI #GradientAccumulation #DeepLearning #LLMs #ModelTraining #NLP #DataScience #Optimization

Understanding Global Average Pooling and Flatten in Large Language Models (LLMs) In the realm of neural networks and LLMs, two key techniques often utilized are GlobalAveragePooling and Flatten. Both play crucial roles in transforming data, but they serve different purposes and offer unique advantages. GlobalAveragePooling - Purpose: Reduces the dimensionality by taking the average of each feature across all timesteps. -Use Case: Particularly useful in sequence data and convolutional layers, it helps in reducing overfitting by summarizing the feature maps. - Example: In sentiment analysis, it can average the learned features from different parts of a sentence, providing a robust representation. Flatten - Purpose: Converts multi-dimensional inputs into a single 1D vector. - Use Case: Commonly used before dense layers to prepare data for classification or regression tasks. - Example: In image processing, Flatten is used to transform the 2D pixel arrays into a format suitable for fully connected layers. Key Differences - Dimensionality Reduction: GlobalAveragePooling reduces dimensions by averaging, preserving essential features and reducing overfitting. - Feature Retention: Flatten retains all features but in a single vector, which can lead to larger model sizes and potentially more overfitting. Enhancing LLM Performance - Regularization Techniques: Dropout and Batch Normalization can be applied to improve generalization. - Advanced Architectures: Using architectures like transformers and incorporating attention mechanisms. - Transfer Learning: Leveraging pre-trained models and fine-tuning them on specific tasks for better performance. #MachineLearning #NLP #DeepLearning #AI #NeuralNetworks #DataScience #TensorFlow #Keras

I'm thrilled to share my recent project on spam detection, where I've leveraged multiple machine learning techniques to enhance accuracy. Initially, I utilized TF-IDF vectorization with Logistic Regression, achieving an impressive 95% accuracy. However, to further experiment, I applied Word2Vec embeddings with both MultinomialNB and GaussianNB, which yielded accuracies of 85.9% and 85.7% respectively. Finally, integrating an SVM model with a linear kernel significantly boosted performance, attaining a remarkable 97.9% accuracy. This exploration underscores the effectiveness of various models and techniques in text classification tasks. Special thanks to #Encryptix for their invaluable support and resources! #MachineLearning #DataScience #SpamDetection #NLP #AI

We concluded an exciting third week at the "AI Solution Architect" course with ELVTR. The combination of lectures and discussions was highly engaging. We delved deeply into the real-world application of NLP, explicitly focusing on text toxicity moderation. Witnessing how NLP can sift through massive datasets to identify and moderate toxic content was fascinating and highlights its role in making digital spaces safer for everyone. The discussions on biases in machine learning were eye-opening. We explored how these biases can affect algorithms and the importance of addressing them to ensure fairness and accuracy in ML systems. Another highlight was the QA process for ML. Training to rigorously test and validate our models ensures they perform well under various scenarios. Furthermore, we started laying out the pipeline deployment for ML projects. Every piece of the model fits into the larger picture of operational deployment. Here's to more breakthroughs and sharing in the coming weeks! #MachineLearning #NLP #AI #DataScience #ELVTR #DucHaba

🚀 Finally published our first paper: "Healthcare Data Summarization via Medical Entity Recognition and Generative AI" ! 💡 We have surpassed the limitations of automation and interpretation of EHRs (Electronic Health Records). How? By creating automatic clinical summaries and intuitive dashboards with LLMs and NLP to enhance medical diagnosis and decision-making. 🔍 Discover more about our approach: https://lnkd.in/duAfKDTh A big thank you to the "Italian Conference on Big Data and Data Science 2023" for this amazing opportunity and our professors! Authors: Antonio Romano Giuseppe Riccio Andriy Korsun Marco Postiglione #healthcare #ai #nlp #llm #datascience #dataengineer