Sagar Das

"Transformers from Large Language Models: A 100-minute lecture" The heart of LLMs like GPTs is transformers. If you have attempted to learn transformers, you must have realized that it is very hard. Each time I try to understand it further, I learn something new. At Vizuara, we wanted to create a comprehensive webinar series on transformers that will serve as the go-to resource for anyone who wishes to learn transformers, including its implementation. We have hosted a 100-minute lecture on our YouTube channel: https://lnkd.in/g2GC938P We have covered the following points: 1️⃣ History of Large Language Models (LLMs) 2️⃣ What are transformers? 3️⃣ Intuition behind transformers 4️⃣ Transformer architecture 5️⃣ NanoGPT 6️⃣ Applications of transformers for LLMs 7️⃣ Building awesome hands-on projects! Here is the link to the lecture The amount of blood, sweat, and research that went into making an almost beginner-friendly workshop was enormous. All credit to Sahil for that. Your work has been very inspirational. We strongly believe that this lecture will tremendously benefit you if you wish to learn about LLMs, transformers, attention mechanisms, the intuition behind positional encoding, etc. in detail.

48

2 Comments

🚀 Bengaluru-based Conversational AI Startup Gnani AI Raises $3M 💰 Gnani AI, India's first voice-first small language model (SLM) for vernacular languages, has secured $3M in funding from tech holding company Info Edge. The investment will fuel the startup's expansion plans. Key Highlights: • Gnani AI is building a series of voice-based SLMs, trained on millions of hours of proprietary audio datasets and billions of conversations in 12 Indian languages • The startup has over 100 customers in banking, financial services, insurance, telecom, automotive, and healthcare industries • Gnani AI's models run on the edge, reducing costs and enabling quick fine-tuning based on enterprise needs • The startup specializes in creating customized models for individual companies, adding real value to their data CEO Ganesh Gopalan emphasizes the importance of multimodal models, focusing on voice and text fusion. By taking their models to enterprises and helping them build on top of it, Gnani AI is revolutionizing the conversational AI landscape in India. #ConversationalAI #VoiceFirst #IndianLanguages #Funding #Expansion

12

1 Comment

Krish Naik recommended this fantastic blog about Transformers, and I stumbled upon it while learning about NLP. The blog breaks down how Transformers function, from attention techniques to their role in speeding up tasks. If you're interested in NLP, this blog is essential reading. Make sure you don't miss it! https://lnkd.in/gDKhwHFy #Knowledgesharing #learningjourney #datascientist #NLP #Transformers #AI #opentowork

26

2 Comments

I'm happy to share that I have recently crossed the 600-citation mark. From multiple paper rejections in the beginning, to serving on the program committees of top NLP conferences like ACL, it was a great experience. Citations from papers published by researchers from - top companies like Google, DeepMind, Microsoft etc. and - top institutes like MIT, Harvard, CMU, etc. are always special. A big thanks to Ajit Rajasekharan (Ajit sir, CTO of nference, Indian Institute of Science (IISc) Almuni) for all his encouragement and support. #research #success #nlproc #deeplearning #llms #generativeai #ai

134

14 Comments

🚨ACL 2024 Outstanding Paper Award Alert 🚨 ➡️Paper Title: INDICLLMSUITE: A BLUEPRINT FOR CREATING PRE-TRAINING AND FINE-TUNING DATASETS FOR INDIAN LANGUAGES 🌟Few pointers from the paper 🎯Despite the considerable advancements in English LLMs, the progress in building comparable models for other languages has been hindered due to the scarcity of tailored resources. 👁In this paper Authors work aims to bridge this divide by introducing an expansive suite of resources specifically designed for the development of Indic LLMs, covering 22 languages, containing a total of 251B tokens and 74.8M instruction-response pairs. 🎯Recognizing the importance of both data quality and quantity, their approach combines highly curated manually verified data, unverified yet valuable data, and synthetic data. 🎯They built a clean, open-source pipeline for curating pre-training data from diverse sources, including websites, PDFs, and videos, incorporating best practices for crawling, cleaning, flagging, and deduplication. 🎯For instruction-fine tuning, they amalgamate existing Indic datasets, translate/transliterate English datasets into Indian languages, and utilize LLaMa2 and Mixtral models to create conversations grounded in articles from Indian Wikipedia and Wikihow. 🎯Additionally, they addressed toxicity alignment by generating toxic prompts for multiple scenarios and then generated non-toxic responses by feeding these toxic prompts to an aligned LLaMa2 model. 🎯 They hope that the datasets, tools, and resources released as a part of this work will not only propel the research and development of Indic LLMs but also establish an open-source blueprint for extending such efforts to other languages. 🏢Organization: 🛎️Nilekani Centre at AI4Bhārat🛎️ Indian Institute of Technology, Madras 🛎️Microsoft🛎️Sant Longowal Institute of Engineering and Technology 🛎️Indian Institute of Information Technology Design & Manufacturing Kancheepuram, 🛎️Maharaja Agrasen Institute Of Technology, Delhi 🛎️Sarvam ,🛎️ NICT 🧙Paper Authors: Mohammed Safi Ur Rahman Khan, Priyam Mehta , Ananth Sankar , Umashankar Kumaravelan, Sumanth Doddapaneni, Suriyaprasaad G, Varun Balan, Sparsh Jain, Anoop Kunchukuttan, Pratyush Kumar, Raj Dabre, Mitesh Khapra 1️⃣Read the Full Paper here: https://lnkd.in/gqwQST-r 2️⃣Code: https://lnkd.in/giWPVntx 🎉 🎉  Congratulations to all the authors 🎉 🎉 Find this Valuable 💎 ? ♻️REPOST and teach your network something new Follow me 👣, Naveen Manwani, for the latest updates on Tech and AI-related news, insightful research papers, and exciting announcements. #ACL2024 #Indicllms #llms #llama2 #mixtral

16

𝗜𝗻 𝘁𝗵𝗶𝘀 𝗮𝗴𝗲 𝗼𝗳 𝗰𝗮𝗽𝘀𝘂𝗹𝗲 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴, 𝗼𝗻𝗹𝘆 𝗱𝗲𝗲𝗽, 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝘄𝗶𝗹𝗹 𝗵𝗲𝗹𝗽 𝘆𝗼𝘂 𝘀𝘁𝗮𝗻𝗱 𝗼𝘂𝘁. If you are serious about transitioning to ML, don't start with toy Kaggle projects like millions. Start by building a strong foundation. 𝗟𝗲𝗮𝗿𝗻 𝘁𝗵𝗲 𝗳𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴:- 𝟭) 𝗟𝗶𝗻𝗲𝗮𝗿 𝗮𝗹𝗴𝗲𝗯𝗿𝗮 𝟮) 𝗣𝗿𝗼𝗯𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝟯) 𝗦𝘁𝗮𝘁𝗶𝘀𝘁𝗶𝗰𝘀 𝟰) 𝗖𝗮𝗹𝗰𝘂𝗹𝘂𝘀 𝟱) 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻 You will not find Machine Learning intimidating. On Vizuara's YouTube channel, I have released an entire course on "𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀 𝗳𝗼𝗿 𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴" where we cover the above topics along with programming fundamentals. If you are a beginner, you can start with this. You will be already ahead of 99.9% of the people. Here are the lectures I have published so far. More will be published on this playlist: https://lnkd.in/g2K97xHW 1) Course introduction: https://lnkd.in/gDdj2Wim 2) Linear Algebra: Vector transformation, span and basis: https://lnkd.in/gBbrSnMd 3) Linear transformation as a matrix multiplication: https://lnkd.in/gpJTNauq 4) Product of two matrices as composite transformation: https://lnkd.in/gWb3RWsD 5) 3D linear transformation: https://lnkd.in/g94VzbeX 6) A simple physical intuition for determinants: https://lnkd.in/g_5JnYbD 7) Transformation with non-square matrices (2D to 3D): https://lnkd.in/gXS3Ggmg 8) Matrix inverses and their physical meaning in transformations: https://lnkd.in/g6pHWj-Q 9) Relation between dot product and transformations: https://lnkd.in/geEPwAbi 10) Simple intuition of eigenvalues and eigenvectors: https://lnkd.in/gg6dVPVM 11) Probability and statistics an introduction: https://lnkd.in/g-Ze6UY5 12) Introduction to conditional probability: https://lnkd.in/gJyr4yFs 13) Intuition and basics of Bayes theorem: https://lnkd.in/gvt4vTQT

163

9 Comments

GPT4o for voice typing ------------------------------------- So far, accuracy with Sellinam, Google voice typing for #tamil is not 100%. After spelled, it requires edit. Tried #GPT4o to output the spoken words in Tamil. It is accurate and doesn't require any hard edit. கேட்பதும் நீ மறுப்பதும் நான் கெஞ்சுவதும் நீ இளகுவதும்... Disclaimer - Haven't tried complex sentences yet. Haven't used any other LLM for voice typing prior.

9

There is a gloom simmering in the data science industry. Confetti is flowing in the wind of generative AI. But at the same time an entire batch of freshers are anxious about their placements. Hiring has been extremely low across the board. Irrespective of tier-1, 2 or 3 campuses. These days whenever any student enquires about our courses, they mainly try to understand whether we provide "placement assistance" or not. Meaning if we will be able to find them a job or not. This is a great indicator to understand the overall mood of the batch. They're sadly, desperate for an opportunity. Looking at the industry from this vantage point breaks you a little bit. So from last two weeks I've been doing research on the topic. Talking to several industry veterans, hundreds of students, reading. And the conclusion is simple and worrying. AI adoption in everyday life is the future. We're moving towards an age where "internet of things" will be a reality pretty soon. Most of the work that can be automated will definitely be automated. As we'll be increasingly a more capitalistic society, most of the necessities will be controlled by companies. And as we all very well know, automation is an effective cost cutting measure. So many people will loose jobs, a lot of people will have to learn new skills to adjust, payouts will follow a leptokurtic distribution. Overall it'll be a gradually more competitive and skill-based economy. Harder to thrive in. Even though it might sound like doom. Life always finds a way. A different way for sure. But definitely a way. Which frankly I don't know yet. I'm planning to read more about this topic. it's a ocean of knowledge to consume. I'm in the water but still near the bay. What you'll see soon is something beautiful, that has come of these research. Till then please Stay Tuned. #lastbench #resumereview #coverletter #interviewpreparation #interviewtips #placement #placementassistqnce #datascience #analytics #machinelearning #careercoach

134

2 Comments

🚀 Unveiling 'Attention Is All You Need': Insights from My Notes and Code Implementation I'm excited to share my comprehensive exploration of the groundbreaking paper "Attention Is All You Need." This paper introduced the Transformer architecture, revolutionizing the approach to sequence modeling tasks in AI. Alongside this, I’ve detailed the inherent challenges with traditional models like RNNs and BiRNNs and compared these with the transformative capabilities of Transformers. Challenges with RNNs and BiRNNs: - Sequential Processing: RNNs process data points one at a time, which significantly slows down training and limits the ability to handle longer sequences effectively. - Vanishing Gradient Problem: As the sequence length increases, RNNs often lose information from earlier points, making it difficult to learn dependencies in longer texts. - Limited Contextual Understanding: While BiRNNs can access past and future context by processing data from both directions, their understanding is still constrained to a fixed window, affecting overall performance. Architectural Overview in My Notes: - RNNs and BiRNNs:I explore how these models are structured to process sequential data and their limitations in depth and context handling. - Transformers: A shift from recurrent processes to a model based solely on attention mechanisms, allowing parallel processing of data and a comprehensive global view of all input tokens simultaneously. What You’ll Find in My Notes and Code: - Detailed Architectural Comparisons: From the basic building blocks of RNNs and BiRNNs to the advanced, multi-headed self-attention and positional encoding strategies of Transformers. - Step-by-Step Code Implementation:Follow along with my code as I implement the Transformer model, illustrating how each component functions and interacts. I am uploading my detailed notes and the full code implementation. These resources will not only help you understand the "why" behind the shift towards Transformers but also the "how" of implementing these models effectively. This deep dive into Transformer technology is designed to clarify the significant advancements over past architectures and to inspire further exploration and innovation in the field. Reference: Code implementation: Famous research paper Attention is All you need! https://lnkd.in/gF6XZni7 #AI #MachineLearning #DeepLearning #NLP #Transformers #AttentionisAllyouneed #selfattention #attentionmechanism

20

Post 1/2: General guidelines for your ML Technical interviews: Before Applying: Research the company and team in advance. Check out their blogs and the prospective team members. Try to align and highlight your skills and key projects that are aligned to the company or the team or the role you are applying to. For example: Let's say you have have completed a variety of ML projects in both computer vision and NLP. If you are applying for an LLM engineer role, you might want to highlight your courseworks, projects, etc. related to NLP. Try to highlight those libraries that you have used for such projects. While writing CVs, try to write concisely in 1-2 lines and use bullet points if possible, instead of writing long sentences and paragraphs. Don’t Inflate your skills and contributions: I have noticed that a few candidates mention their group projects, without having core contributions and passing it as their own, while their contributions are almost marginal. During interviews, when asked about such group projects and their details, the candidates bail out saying they don’t know because they were not working on that particular aspect and they had worked on other aspects. This may bring out a bad impression to your interviewer. Now, the question is should you include or not such projects in your CV? Personally, I think it is fine to add such projects, as long as you have good understanding and you are able to explain the underlying concepts and all the aspects of the projects, while clearly mentioning your contributions. Putting lots of jargon and keywords: Your resume might get noticed in the initial stages, when someone might be searching for potential candidates and the keywords in your resume might get you lucky. However, on the long run they might hurt The more keywords you add and highlight, there is a high chance the interviewer may assume that you have thoroughly understood or applied those concepts. Unless you are prepared to explain those concepts in detail along with their intuition and understanding, I would suggest not using unnecessary jargon. Continue reading in Post 2/2... https://lnkd.in/gQGjMGJ9 P.S. This post does not generalize the interview process and varies from interviewer to interviewer. Each organization may have their own interview policies and may have their own style of interview process. #machinelearning #interviews #interviewpreparation #interviewtips

60

2 Comments

Sarvam AI Unveils Sarvam-1: A Game-Changer in Indian AI! India’s Sarvam AI, co-founded by Vivek Raghavan and Pratyush Kumar, is headquartered in Bangalore and has launched Sarvam-1, a cutting-edge language model designed specifically for Indic languages. And it’s already outperforming the big players like Gemma-2 and Llama-3.2! Here’s why Sarvam-1 is making waves: ✅ Supports 10 Indian Languages + English: From Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Punjabi, Gujarati, Oriya, to English—this model bridges the language gap in AI-powered communication. ✅ Outshines Larger Models: With only 2 billion parameters, Sarvam-1 excels across benchmarks like ARC-Challenge and IndicGenBench. On TriviaQA, it achieved 86.11 accuracy across Indic languages, towering over Llama-3.1 8B’s score of 61.47. For translation, summarization, and cross-lingual tasks, Sarvam-1’s chrF++ score of 46.81 on the Flores dataset is unmatched. ✅ 4-6x Faster Inference: Its speed makes it ideal for edge devices with limited computing power. ✅ Efficient Tokenization for Indian Scripts: Traditional LLMs require more tokens per word for Indian languages, but Sarvam-1’s tokenizer brings this down to 1.4–2.1 tokens, similar to English’s efficiency. ✅ Trained on the Sarvam-2T Dataset: This 2-trillion-token corpus contains a rich blend of technical, scientific, and programming data across the 10 languages, enhancing its ability to handle complex reasoning tasks. Built on Yotta’s Shakti cluster with NVIDIA’s NeMo framework, the model was trained on 1,024 GPUs in just five days, showing incredible efficiency. Developers can now explore Sarvam-1 on Hugging Face, opening new doors for Indic NLP applications—from conversational AI to advanced translations. With Vivek Raghavan and Pratyush Kumar leading the charge, Sarvam AI is at the forefront of AI innovation in India, setting new standards in multilingual technology. The future of AI-powered conversations is here, and it’s proudly made in Bangalore! I am eagerly waiting for Sarvam's API to test it and develop meaningful use cases. Are you as well? Please feel free to comment below!" #AI #SarvamAI #MultilingualAI #TechForIndia #HuggingFace

37

30 sec analysis of some Kannada Instruct Tuning Dataset on Hugging Face. 1. Who writes code in Kannada😒? 2. Its contamination of LLM weights if you instruct LORA backprop on such garbage data without using logic, leading to more catastrophic forgetting. 3. थोड़ी मर्यादा रखिए 🙏। [not pasting link coz they are startup] If anyone thinks differently can add valuable insights into comments.

6

🚨Paper Alert 🚨 ➡️Paper Title: Loong: Generating Minute-level Long Videos with Autoregressive Language Models 🌟Few pointers from the paper 🎯It is desirable but challenging to generate content-rich long videos in the scale of minutes. Autoregressive large language models (LLMs) have achieved great success in generating coherent and long sequences of tokens in the domain of natural language processing, while the exploration of autoregressive LLMs for video generation is limited to generating short videos of several seconds. 🎯In this work, authors conducted a deep analysis of the challenges that prevent autoregressive LLM-based video generators from generating long videos. Based on the observations and analysis, authors have proposed “Loong”, a new autoregressive LLM-based video generator that can generate minute-long videos. 🎯Specifically, they modeled the text tokens and video tokens as a unified sequence for autoregressive LLMs and trained the model from scratch. They proposed progressive short-to-long training with a loss re-weighting scheme to mitigate the loss imbalance problem for long video training. 🎯They further investigated inference strategies, including video token re-encoding and sampling strategies, to diminish error accumulation during inference. 🎯Their proposed Loong can be trained on 10-second videos and be extended to generate minute-level long videos conditioned on text prompts, as demonstrated by the results. 🏢Organization: The University of Hong Kong , ByteDance 🧙Paper Authors: Yuqing Wang, Tianwei Xiong, Zhou D., Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng , Xihui Liu 1️⃣Read the Full Paper here: https://lnkd.in/gcE_-5wF 2️⃣Project Page: https://lnkd.in/gU3rU6vU 🎥 Be sure to watch the attached Demo Video - Sound on 🔊🔊 Find this Valuable 💎 ? ♻️REPOST and teach your network something new Follow me 👣, Naveen Manwani, for the latest updates on Tech and AI-related news, insightful research papers, and exciting announcements.

9

DL interview question: What is Dropout? Dropout is a technique used in deep learning to prevent overfitting. Overfitting occurs when a model learns too much details(ie noise, irrelavnat patterns rather than generic patterns) from the training data and fails to perform well on new, unseen data. How does dropout work? 1. Randomly turns off neurons: During training, dropout randomly sets a percentage of neurons (e.g., 50%) to zero. These neurons do not contribute to the model’s learning in that iteration. 2. Creates different sub-networks: Each training step effectively trains a slightly different model since different neurons are turned off. This helps the model learn diverse patterns. 3. Adjusts for Predictions: When making predictions (inference), all neurons are used, but their outputs are scaled down to ensure consistency with training. Why is dropout important? 1. Reduces overfitting: By not relying on specific neurons, dropout helps the model learn more generalized features that work better on new data. 2. Improves performance: Models that use dropout typically perform better on unseen data, making them more reliable in real-world applications. 3. Increases robustness: Dropout helps the model adapt to noise and variations in the data, leading to more stable performance across different inputs. 4. Encourages feature learning: It promotes learning multiple features rather than focusing on a few, which can enhance overall model performance. 5. Easy to implement: Dropout is straightforward to apply and adds minimal extra computation, making it a popular choice in many neural networks. #Deeplearning #Overfitting #Dropout #InterviewQuestion

358

8 Comments

DSA is not required for Data scientist? Let us take one of the ground breaking approaches in NLP as our reference. -> RAG (Retrieval Augmented Generation) [Check out my previous post if you are not familiar or want revision https://lnkd.in/gVwzsnQk] One of the most popular implementations of RAG utilises HNSW (Hierarchal Navigable Small Worlds) Graphs and Heaps are the major component of it's functioning -> Hierarchical Structure: The algorithm organizes the data points into multiple layers of graphs. The top layers are sparse and contain nodes with long-range connections (fewer, more global relationships between points). The lower layers are denser and contain more local, nearest neighbor connections. -> Navigating the Small World: The search process begins at the topmost layer and proceeds downward. The algorithm moves from higher layers (global, sparse connections) to lower layers (local, dense connections) to refine the search. At each layer, the search starts from a random entry point and navigates the graph using greedy search to find the closest neighbors based on the distance metric (e.g., Euclidean distance). -> Multi-layer Traversal: When moving down the layers, the search focuses more on local neighbourhoods to ensure accuracy in finding the nearest neighbours. So alongside performance metrics, hyper tuning and other relevant approaches, building low level algorithms will definitely help strengthen your knowledge based. For mentorship: https://lnkd.in/gm5GfqpD Discount code: ROHAN80

21

1 Comment

🚀 Exciting Progress on Real Time Translation Project! 🎤 Over the past couple of days, I have been experimenting with the English Multispeaker Corpus for Voice Cloning dataset provided by Kaggle using Google Colab due to its flexibility and deals with warnings & errors much better than Kaggle notebooks . 💻🔧 I have been splitting the dataset into training and testing sets, trying to refine my feature selection and label identification. 📊 I got an amazing 99% accuracy for testing individual speakers, but the results have some overfitting issues causing it to fail all together. Which means I needed to dig deeper in feature extraction and researching how I can avoid these issues. 🔍📈 To enhance my understanding, I have decided to start watching various channels that have done remarkable work on how to work with audio data. I have been taking notes on how I can integrate these techniques such as CNNs or RNNs and much more to improve the efficiency in my project. 📚🎶 Also recently, I have identified that only 21 audio files are common amongst all speakers in the date set which was an exciting finding because it will help me streamline my testing and training processes! 🎉 I have spent a responsible amount of time to organize the data, create loading functions, and plot wave forms for visual analysis. 📈🎧 But currently I am creating a preprocessing function that will generate the spectrograms, which I'll use as input for a CNN model for classifying speakers. 🖼️🤖 Also shoutout too TensorFlow it has been such a game-changer for me and I am so excited for how far me and my team can go to really make this project outstanding and hopefully useable for all of us! 🚀🔍 LINKS TO VIDEOS THAT I AM CURRENLY USING: 🎧 Build a Deep Audio Classifier with Python and Tensorflow: https://lnkd.in/errvymFp EDA-Audio Classification Project Using Deep Learning: https://lnkd.in/e3rY5ndR Deep Learning for Audio Classification: https://lnkd.in/eE775icZ Stay tuned for more updates! 💡🙌 #VoiceCloning #DeepLearning #TensorFlow #CNN #AudioProcessing #MachineLearning #DataScience

8

Balancing perfomance and cost is one of the key factors in deploying LLMs in prod, we often use multiple LLMs for load balancing/routing the queries to handle the TPS(Tokens Per Second)/RPS(Requests Per Second). Recently came over this approach of RouteLLM by LMSys where the authors have trained a multiple router models using different approaches such as using Matrix Factorization methods snd causal LM approaches to effectively reduce the cost of the inference by routing "complex" questions to strong/large model(such as GPT4) and "simple" questions to weak/small model(such as Llama) The framework uses human preference data from Chatbot Arena along with data augmentation to train multiple routers which reduces the cost upto 70% cost reduction in MT Bench. Interestingly this approaches are transferable to other models as well. Would recommend this paper[1] a read and also the codebase[2] [1] Paper: https://lnkd.in/gSHfdMEw [2] Github: https://lnkd.in/gDmvpGVj #machinelearning #generativeai #llms

45

🚀 𝐈𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐢𝐧𝐠 𝐋𝐥𝐚𝐦𝐚3 𝐟𝐫𝐨𝐦 𝐬𝐜𝐫𝐚𝐭𝐜𝐡! 🚀 🔷Covers every detail from the tokenizer to attention mechanisms 🔷Includes embeddings, multi-head attention matrix multiplications, and positional encoding 🔷Complete implementation of all intermediate layers Checkout 📰 𝘓𝘪𝘯𝘬𝘴 𝘪𝘯 𝘤𝘰𝘮𝘮𝘦𝘯𝘵 𝘴𝘦𝘤𝘵𝘪𝘰𝘯 👇 Follow Girraj Jangid ✨ for more insights. Thanks, Nishant Aklecha #llama #meta #llms #ai

10

1 Comment

By focusing on different points of the conditional distribution, we can identify patterns and relationships that might go unnoticed with mean regression techniques. This is particularly advantageous in real-world applications where the mean does not always offer a complete picture. For instance, in economics and finance, understanding the behavior of the lower or upper tails of the distribution can be crucial for risk management and policy making. Our journey into quantile regression begins with a foundational understanding of quantiles themselves, which are essentially cutoff points dividing the range of a probability distribution into contiguous intervals with equal probabilities. From there, we delve into the methodology of quantile regression, which extends the concept of quantiles to the conditional distribution of the response variable, hence allowing for a more detailed exploration beyond the central tendency. Victor Chernozhukov and Ivan Fernandez-Val. 14.382 Econometrics. Spring 2017. Massachusetts Institute of Technology: MIT OpenCourseWare https://lnkd.in/egT_Qpt5

7

1 Comment