Faruk Ahmad’s Post

Two techniques — Matryoshka Representation Learning and Binary Quantization Learning — unlock new possibilities, making the once-impossible possible. Read our latest article on The New Stack!   

🚀 Adaptive AI Learning 🚀 It is important to learn new tasks without forgetting the old ones, EASE offers a smart approach with dedicated adapters for each new topic. This ensures continuous learning for AI systems. Paper 🔗 - https://lnkd.in/grRW2GDd 💡 Keen on AI breakthroughs? DM us to explore opportunities!

Interpretable Features in LLMs

Hello connections , Excited to share my latest project: LSTM Text Generation using TensorFlow! 🎉As part of my journey into deep learning and natural language processing (NLP), I recently worked on a project focused on building a text generation model using Long Short-Term Memory (LSTM) networks with TensorFlow. The project involved constructing an LSTM-based architecture to generate coherent and contextually relevant text based on sequential data. To prepare for this, I deep-dived into LSTM and RNN concepts, drawing significant insights from the influential paper An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling. This paper provided valuable theoretical foundations, especially regarding how LSTM models can capture long-term dependencies in sequences, which was crucial to the success of my project.The project began with data preprocessing, where I worked on tokenizing and vectorizing text data to make it suitable for model training. After that, I focused on model design, training, and fine-tuning the network to achieve creative and coherent text outputs. It was an enriching experience to see how neural networks could grasp the structure of text and produce meaningful sequences.This hands-on experience has deepened my understanding of sequence models and the power of NLP. I’m excited to explore more in this space and apply these learnings to future AI challenges. This project was a great opportunity to apply advanced techniques in AI and enhance my skills in TensorFlow. I’m grateful for the support and resources provided by my mentors and peers throughout this journey. A big thank you to my mentors Nagendra Kishore Girajala sir, Aravind Pappala sir, for their valuable guidance and feedback. I’m eager to continue exploring new challenges and innovations in the AI field! A special thank you to Babji Neelam, the CEO of Technical Hub , for the incredible opportunity to work on AI . I am excited to contribute my skills and insights to advance our AI capabilities and make a meaningful impact with this innovative work. https://lnkd.in/gT7KDFxg #MachineLearning #DeepLearning #NLP #AI #LSTM #RNN #TextGeneration #TensorFlow #DataScience #ArtificialIntelligence #NeuralNetworks #SequenceModeling #NaturalLanguageProcessing #AIResearch #DLFrameworks #MLProjects

Here is part II of my LLM lecture: https://lnkd.in/e9yqfCDM 0:00 Fine-tuning LLMs 4:58 Low Rank Adaptation (LORA) 21:26 Quantization 41:06 QLORA 46:43 Prefix Tuning 52:22 Retrieval Augmented Generation (RAG) 1:06:26 In-context Learning 1:18:45 Chain-of-Thought Prompting

Soheil Feizi is excellent, and his course is invaluable for covering fundamental concepts in deep learning to the most state-of-the-art, all in one go.

InconsistencyMasks: A TensorFlow Breakthrough in Deep Learning for Image Segmentation-Michael Vorndran InconsistencyMasks is a TensorFlow User Group (TFUG) implementation that introduces a novel method for image segmentation in the field of deep learning. It addresses the challenge of generating sufficient labeled data, especially in uncharted territories of image segmentation. The project introduces the use of Inconsistency Masks (IM) to effectively filter uncertainty in image-pseudo-label pairs, thereby improving segmentation quality. The approach has been tested on the ISIC 2018 dataset and other datasets, consistently achieving exceptional results. It also includes an extensive analysis of prevalent semi-supervised learning strategies. #InconsistencyMasks #TensorFlow #DeepLearning #ImageSegmentation #AIInnovation #SemanticSegmentation #SemiSupervisedLearning #MachineLearning #ResearchInAI #TensorFlowImplementation #InnovationInDL #NeuralNetworks

🚀 New technique unlocks significant performance gains for Large Language Models (LLMs) Just came across an intriguing paper by Matteo Pagliardini, Pierre Ablin, and David Grangier on the state-of-the-art AdEMAMix optimizer. Feeling inspired, I took my keyboard and implemented it myself using PyTorch. You can check out implementation here: https://lnkd.in/d9q2BN7Q During my initial experiments, I noticed improvements compared to the baseline optimizer AdamW. Stay tuned for a blog post where I explore deeper. You can find original paper in the comments section. #AI #MachineLearning #Optimization #PyTorch #LLM #LargeLanguageModels #DeepLearning #Innovation

In a detailed, hands-on guide to diffusion models, Nick DiSalvo walks us through a full implementation of a denoising diffusion probabilistic model (DDPM) in PyTorch.

From Inference Scaling to Problem Graphs: A New Approach to Complex Question Answering with LLMs Reading Inference Scaling for Long-Context Retrieval Augmented Generation sparked an idea: what if we use a Problem Graph approach for handling complex, multi-hop questions? Instead of relying solely on iterative retrieval, LLMs could map out a question’s structure by generating a graph where each node is a sub-question. Inspired by RAG’s retrieval strategies, this method allows the model to explore paths step-by-step and retrieve information strategically. Setting limits on graph exploration prevents unnecessary branching, while summarizing the entire graph at the end delivers a well-rounded answer. This approach, blending RAG insights with graph exploration, could make solving complex questions both efficient and insightful! https://lnkd.in/edf3x2sm