Mehak Beri

In my early learning phase, I often confused Layer Normalization and Batch Normalization—I knew they improved numerical stability, but the key differences eluded me. Watching Umar Jamil's videos clarified everything: Batch Normalization works across the batch dimension, ideal for CNNs, while Layer Normalization focuses on features, perfect for RNNs and Transformers. Grateful for such insightful resources! 🙌 #deeplearning #machineLearning #GENAI #Transformer

10

❝ How do you taste your Wine? With the right Glass, served from the right Bottle. How do you utilize your LLM? With the right Task, served from the right Compute environment. ❞ If the quote above caught your interest, check out my blog post that was featured on the Toyota Connected India Medium channel last week: Leveraging LLMs in Cloud - 🔗 https://lnkd.in/gfHsdevH --- I hope this blog adds a bit of amusement for the seasoned AI/ML engineers 🧠👨‍💻 with the carefully crafted images and perhaps even the silly wine analogies 😉. Whether you're a cloud engineer ☁️ 👨‍💻, a product owner of cloud-based apps, or someone new to Generative AI in the cloud, this blog offers a detailed, beginner-friendly look into the selection and deployment of LLMs in cloud environments. --- A heartfelt thanks 🙏 to my colleagues at Toyota Connected India for taking the time to review, shape this piece and even share a creative writeup like below in the original post 🫡 👇 🤔 Curious to dive in? Click the link below to explore the blog on Medium! 🌐 🔗 https://lnkd.in/gfHsdevH

18

Excited to share our interactive BlogApp on Generalized Linear Models (GLMs)! In collaboration with Venkatachalam Subramanian Periya Subbu, Satya Lasya N, and Pooja Baralu Umesh, we've developed an intuitive tool that makes understanding GLMs both engaging and accessible. If you are an aspiring Data Scientist, Actuary, Statistician, Economist, Epidemiologists or an Engineer, you definitely need to check it out at: https://lnkd.in/gMqf3M5Y View the source code: https://lnkd.in/g9HUm_7r Special thanks to Professor Cody Carroll for the guidance throughout this project. Let's discuss GLMs! What's your experience with these powerful statistical tools? Share your thoughts below. #DataScience #Statistics #MachineLearning #GLM #Analytics #Programming #USF #MSDS

34

2 Comments

Just finished RAG and Fine-Tuning Explained! Check it out: https://lnkd.in/gbf5cQ4B #generativeai #aisoftwaredevelopment #retrievalaugmentedgeneration

8

Ran a quick test on the new #OpenAI #o1mini to solve IIT-JEE Advanced 2024 maths questions. It accurately solved 13 out of 17 questions with an impressive accuracy of ~76%. Quick thoughts: 1. The hidden chain of thought is quite detailed and extensive. Breaks the problem, solves and verifies and iterates multiple times. This might increase the output token cost by a big margin. 2. Seems there are tools integrated for calculations. Much needed as #gpt4o does a bad job at it. 3. Latency is quite high. Some questions took more than 50 seconds to get answered. This needs to come down for production user-facing use cases. 4. Hallucination still exists. Check the 1st question in the link, it had the correct step-by-step reasoning but hallucinated at generating the final answer. 5. Overdoing self-thought. Saw a couple of instances where it did generate correct answers in the process but eventually made wrong answers. Would give a clearer picture once the rate limits are increased and benchmarking is done on a bigger sample. Let me know any interesting findings and thoughts 😀 https://lnkd.in/giSCezyG #GenAI #OpenAI #PhysicsWallah

139

5 Comments

Join me and Farris in Neurips workshop on Fine-Tuning in Modern Machine Learning: Principles and Scalability. Our paper title is: Optimizing Small Language Models for In-Vehicle Function-Calling

25

I read a paper on ViT-22B by Google Research and here's a summary. Title: Scaling Vision Transformers to 22 Billion Parameters Links: 1. https://lnkd.in/gJa8PsvJ 2. https://lnkd.in/gPiGxKTJ Motivation: Unlike in LLMs, scaling the standard transformer architecture for vision tasks leads to pathological instability like divergent training loss while training the model. The paper demonstrates architectural changes to address such challenges. Architecture Key Points: 1. Parallel layers: Instead of sequential execution as in the standard Transformer, ViT-22B parallelly executes Attention and MLP Blocks, which reduces training time. 2. QK Normalization: To solve the instability of divergent training loss, LayerNorm is applied to the queries and keys before the dot-product attention computation. 3. Omitting biases on QKV projections and LayerNorms: The bias terms were removed from the QKV projections and all LayerNorms are applied without bias and centering. This improves accelerator utilization without quality degradation.

13

I would like to share my first blog post on my personal website: Processing 1 billion rows of data with native Python 3.12. Check it out! https://lnkd.in/g9RY7Gmz

12

2 Comments

On Day 16 of #30DaysOfFLCode, I delved into Federated Transfer Learning (FTL). What is FTL? Combines Federated Learning and Transfer Learning to train models on heterogeneous data from multiple domains. Helps when datasets from different devices or organizations have limited overlap. Applications: Personalized healthcare recommendations using data from hospitals with varying specializations. Cross-lingual AI models where each dataset is in a different language. FTL unlocks new possibilities for collaborative AI across diverse datasets.

8

Excited to share insights from my recent conversation with Shaan Mukhtar and Scott Cecil about the open-source library RAY (https://lnkd.in/gE6sShCe) ⚡️ This discussion opened my eyes to the potential of Ray for accelerating data science workflows. Here's what I learned: - Ray simplifies data ingestion, feature engineering, and deployment, especially on unstructured data. - Ray Serve makes deployment a breeze with just one line of code. - Scaling up and down is effortless, utilizing both CPUs and GPUs. - KubeRay enables seamless integration with Kubernetes for scaling. - Ray is a game-changer for LLMs, providing efficiency and lower latency. - While SQL remains essential, Ray can complement and even replace Spark in certain scenarios like unstructured data. I'm eager to explore Ray further and put these learnings into practice. Thanks to Shaan and Scott for the insightful conversation and personalized connection! 🙏 #Ray #DataScience #MachineLearning #LLMs #Deployment #Scaling

27

4 Comments

🚀 Thrilled to share that I’ve completed the Open-Source Models with Hugging Face course by DeepLearning.AI! This has been an incredible learning journey, exploring a wide range of cutting-edge AI technologies. Throughout the course, I had the opportunity to dive into topics such as: Natural Language Processing (NLP) for text analysis Translation and Summarization tools Sentence Embeddings for capturing semantic meaning Zero-Shot Audio Classification Automatic Speech Recognition and Text-to-Speech Object Detection, Image Segmentation, and Retrieval Image Captioning and Visual Question Answering Zero-Shot Image Classification Deployment strategies for real-world applications Understanding these concepts and tools has been instrumental in gaining insights into building intelligent, real-world applications. Hugging Face models make it easier to integrate state-of-the-art AI into a wide range of projects, from natural language to computer vision and beyond. Excited to put this knowledge to use and explore new possibilities in the world of open-source AI! #AI #MachineLearning #DeepLearning #NLP #ComputerVision #HuggingFace #ArtificialIntelligence

15

Ever wondered how random numbers work in NumPy and why setting a random seed is important? 🌟 Check out my latest blog to learn more about true vs. pseudo-randomness and the role of random seeds. #Randomness #PseudoRandomness #NumPy #PythonProgramming #Algorithms #Computing #Cryptography #ScientificComputing

21

6 Comments

Function calling in Large Language Models (LLMs) is redefining how AI integrates with real-world systems. By enabling models to invoke external functions or APIs, this feature goes beyond generating text to seamlessly performing actions like fetching live data, triggering workflows, or executing specific tasks. For instance, imagine asking an AI to calculate complex formulas, retrieve weather updates, or automate scheduling—it’s all possible with function calling. This capability allows LLMs to generate structured outputs (e.g., JSON) that external systems can interpret and act upon, bridging the gap between conversational AI and actionable outcomes. This evolution makes LLMs not just conversational agents but problem-solving tools for real-time scenarios. What’s most exciting is the potential to build smarter, more interactive applications across industries.

6

2 Comments

recently been tinkering with running hand rolled neural nets on FPGAs for audio generation. ( it's crazy being able to run inference almost 200,000 inferences / second e.g. https://lnkd.in/gqepf6Cy ) stumbled today onto a cool idea called LUTnet ( https://lnkd.in/gi9GiH2Y ) where binary neural nets can be expressed as 2-LUTs ( generalisable to N-LUTs ). what a great idea! has anyone worked with these kinds of architectures? am particularly interested in how i can use the same ideas with my power-of-two quantisation networks (where "multiplies" are just left/right shifts )

26

For the past few weeks, I’ve been exploring JAX and Flax, and I’m really appreciating their design philosophy. A key decision they’ve made is to avoid “reinventing the wheel.” Rather than duplicating functions like dataset loading or checkpointing, JAX/Flax taps into the strengths of PyTorch and TensorFlow, which already offer robust solutions. With a few adjustments, you can seamlessly use these libraries within the JAX/Flax ecosystem for data handling, saving both time and resources. For checkpointing, Flax initially had its own system with `flax.training.checkpoints`, but this has since been replaced by Orbax, which enhances model persistence. However, recent changes in JAX's random number generation (RNG) format have introduced some challenges. The new random key format `dtype=key<fry>` is fully compatible with Flax itself, but Orbax has yet to adopt it, which means that saving checkpoints with dropout layers now requires additional manual adjustments to the key type, adding a layer of complexity. This experience highlights a broader challenge for Flax: its reliance on external libraries that may not always keep pace with its own updates. While the Flax community is actively innovating, dependency on other libraries means they’re sometimes at the mercy of external update cycles. It’s a reminder that in fast-evolving ecosystems, dependencies can sometimes introduce bottlenecks that are outside the control of the primary library's community. Despite these challenges, it’s exciting to see how JAX and Flax continue to push the boundaries, even if it means navigating some temporary friction along the way. #MachineLearning #DeepLearning #JAX #Flax #DataScience #TensorFlow #PyTorch #DeepLearning

12

Day 5 of learning in public 🚀 Traditional fine-tuning methods for large language models often face challenges like high computational requirements, catastrophic forgetting, and single-task specialization. Enter LoRA (Low-Rank Adaptation), a game-changing technique that addresses these limitations, enabling efficient and practical fine-tuning of large language models. LoRA introduces a novel approach to fine-tuning by adding a small number of task-specific weights to the pre-trained model instead of updating all the weights. Here's how it works: 1. Low-Rank Matrices: For each layer in the pre-trained model, LoRA introduces a pair of low-rank matrices: one for the weights and one for the biases. 2. Adaptation: During fine-tuning, LoRA updates these low-rank matrices instead of modifying the pre-trained weights directly. 3. Composition: At inference time, the low-rank matrices are composed with the pre-trained weights to produce the final, adapted weights for the fine-tuned model. The low-rank nature of these matrices is the key to LoRA's efficiency, dramatically reducing the memory and computational requirements for fine-tuning. This makes it possible to adapt large language models even on consumer-grade hardware like CPUs or edge devices. But LoRA's advantages don't stop there. It can match or even outperform the performance of full fine-tuning while using far fewer updatable parameters. Additionally, the LoRA updates are reversible, allowing the model to revert to its original state, and composable, enabling a single pre-trained model to be fine-tuned for multiple tasks simultaneously. To illustrate the power of LoRA, let's dive into a practical example. Here, I'll demonstrate how to implement LoRA from scratch and fine-tune the DistilBERT model using LoRA on the IMDB dataset for sentiment analysis. https://lnkd.in/g-Tz2NU7 As you can see, LoRA is a remarkable technique that addresses the limitations of traditional fine-tuning methods, making it more efficient, practical, and versatile. By introducing low-rank, task-specific updates, LoRA enables efficient adaptation while preserving the model's original knowledge and allowing for multi-task specialization. As the demand for deploying large language models on resource-constrained devices continues to grow, techniques like LoRA will make these powerful models more accessible and practical.

4

🌿 SandalQuest: Digitizing Knowledge of Sandalwood Cultivation with AI 🌿 I’m excited to share our project SandalQuest, created as part of the ML-Fiesta: AI/ML Hackathon organized by International Institute of Information Technology Bangalore! Our team developed an AI-powered pipeline to preserve and provide easy access to the indigenous knowledge surrounding sandalwood cultivation in Karnataka. 🔑 What We Built: - An Automatic Speech Recognition (ASR) model tailored for colloquial Kannada, capturing informal language and dialect variations. - A speech-based question-answering system that allows users to ask questions and retrieve relevant answers from a transcribed audio database. - A searchable audio database that helps users access traditional knowledge through speech. 💻 Tech Stack: - Python, PyTorch for AI & Machine Learning - Whisper for Speech Recognition - Hugging Face Transformers for NLP - MongoDB for database management 🔍 Impact: We’re helping preserve valuable knowledge on sandalwood cultivation and making it easily accessible for farmers, researchers, and enthusiasts. 🔗 Check out the project: https://lnkd.in/dMxRP76s 🚀 Team Name: Code Wizards Team Members: - Shani Sinojiya (AI/ML & Backend Developer) - Mohammad Anas Africawala (AI/ML Engineer) - Tisha Patel (Full Stack Developer) It’s been an amazing experience working on this project, and I’m proud of what we’ve achieved. Looking forward to applying these insights to real-world applications! #AI #MachineLearning #SpeechRecognition #Kannada #ColloquialSpeech #Sandalwood #DataScience #NaturalLanguageProcessing #AIForGood #Hackathon #TechForGood #GoogleCloud #AIResearch #DeepLearning

10

2 Comments