Omkar Singh Rajpurohit

Optimizing ML models in production is crucial for high performance. Techniques like parallelism, model replication, quantization, and image decoding optimization enhance efficiency. By leveraging tensor and pipeline parallelism, models can achieve faster inference while balancing latency and throughput. Model replication scales services across devices, optimizing GPU usage. Quantization reduces model size, improving computational speed. Implementing these strategies ensures optimal performance and scalability in real-world environments, driving efficient ML model serving. Link : https://lnkd.in/gZSjyhja

1

Title: Model-Free Change Point Detection for Mixing Processes Authors: Hao Chen; Abhishek Gupta; Yin Sun; Ness Shroff Date: 8 May 2024 Link: https://lnkd.in/gnyQtgCe This paper studies the problem of change point detection under the completely data-driven setting . A testing procedure named MMD-CUSUM is proposed and analyzed. The proposed method does not require explicit knowledge of the pre- and post-change distribution or the data generating process. The theoretical performance of the MMD-CUSUM statistic is obtained under exponentially alpha, beta mixing and fast phi mixing processes. This significantly extends the current performance analysis of MMD-CUSUM under i.i.d. and Markovian samples, which allows the method to adapt to a wide range of real-world settings without modification. #changepoint #kernelmethod #mixingprocesses

2

📝 A Curated Set of NLP Primers: Attention, Autoregressive vs. Autoencoder Models, Transformers, BERT 🔹 Attention: http://attention.aman.ai - Overview (The Attention Mechanism, The Bottleneck Problem, The Context Vector Bottleneck, How Attention Solves the Bottleneck Problem) - Origins of Attention - Attention: Under the Hood - Classic Sequence-to-Sequence Model - Sequence-to-Sequence Model with Attention - Context Vector - Compared: Attention Mechanism v/s Fixed-length Context Vector Approach - Extensions to the Classic Attention Mechanism - Self-Attention (Why Have Multiple Attention Layers?) - Comparative Analysis: Additive vs. Scaled Dot-Product Attention (Computational Efficiency, Theoretical Complexity, Usage and Performance, Implementation Details) - Multi-head Attention - Cross Attention - Ghost Attention - Linear Attention 🔹 Autoregressive vs. Autoencoder Models: http://autoregenc.aman.ai - How do Decoder Models Work? - How do Encoder Models Work? - Encoder-Decoder Models - XLNet: the Best of Both Worlds (Permutation Language Modeling) 🔹 Transformers: http://transformer.aman.ai - Mathematical background (Vectors, Matrix Multiplication, Dot Product, Masking, Sampling) - Attention (Additive/Multiplicative Attention, Self/Cross-Attention, Multi-Head Attention) - Core components of the Transformer Architecture (Embeddings, Positional Encoding, Skip Connections, Layer Normalization, Softmax) - Top-level Transformer Architecture (Encoder and Decoder stack) - Implementation details (Byte-Pair Encoding, Teacher Forcing, Label Smoothing) - Lessons learned (What are Transformers learning? Why is training them so hard?) - Pros/cons of Transformers relative to CNNs/RNNs - Relation between Transformers and Graph Neural Networks 🔹 BERT: http://bert.aman.ai - Background: Pre-Training - Contextualized Embeddings - Masked Language Modeling (MLM) - Next Sentence Prediction (NSP) - BERT’s Encoder Architecture vs. Other Decoder Architectures - The Strength of Bidirectionality - Supervised Finetuning #artificialintelligence #deeplearning #nlp

52

4 Comments

📊 Check out Ahmed Alrasheed's st-link-analysis component for visualizing graph data using Cytoscape.js! You can customize styles, labels, colors, captions, and icons. 🎨 👉 Learn more: https://buff.ly/3zKDxBs 👀 Demo: https://buff.ly/3zL75il #dataviz #python #datascience

135

3 Comments

Combining Neo4j #KnowledgeGraphs, native vector search, and #Cypher LangChain templates using LangChain agents for dynamic query handling and enhanced information retrieval. Look at this project by Saurav Joshi, which aims to improve the efficiency and accuracy of AI-driven data retrieval and processing. 🔥 🔥 https://bit.ly/4e3NOZg #Langchain #Neo4j #vector #AI

61

⚔️Post 1 🚨Paper Alert 🚨 ➡️Paper Title: VASA-1: Lifelike Audio-Driven Talking Faces Generated in Real Time 🌟Few pointers from the paper 🎯In this paper authors have introduced VASA, a framework for generating lifelike talking faces with appealing visual affective skills (VAS) given a single static image and a speech audio clip. 🎯Their premiere model, VASA-1, is capable of not only producing lip movements that are exquisitely synchronized with the audio, but also capturing a large spectrum of facial nuances and natural head motions that contribute to the perception of authenticity and liveliness. 🎯 The core innovations include a holistic facial dynamics and head movement generation model that works in a face latent space, and the development of such an expressive and disentangled face latent space using videos. 🎯Through extensive experiments including evaluation on a set of new metrics, authors  showed that their  method significantly outperforms previous methods along various dimensions comprehensively. 🎯Their method not only delivers high video quality with realistic facial and head dynamics but also supports the online generation of 512x512 videos at up to 40 FPS with negligible starting latency. It paves the way for real-time engagements with lifelike avatars that emulate human conversational behaviors. 🏢Organization: Microsoft Research Asia 🧙Paper Authors: @Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, aining Guo 1️⃣Read the Full Paper here: https://lnkd.in/ggDQ-dHP 2️⃣Project Page: https://lnkd.in/gCgUjWV4 🎥 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. More Mind Boggling example from VASA-1coming in next post Stay Tuned 💺💺 #avatars #lipsync #microsoft

7

"Graph Neural Network from Scratch" There is a lot of theory material available for learning GNN. But very few sources from where you can learn to implement Graph Neural Networks using code. Our ML project-based course on GNN on Vizuara’s YouTube channel intends to fill this gap by implementing a Graph Neural Network from scratch. In the course instructor Aiswarya’s latest lecture, she demonstrates how we can implement a graph neural network on a dataset of a set of students studying at a university, and predict which students are likely from the same country of origin. Here is the lecture https://lnkd.in/gFZfMwsi If you have spent some time on GNNs, you may have the following questions. 1) If we do not assume any features, how do we perform one-hot encoding to create the feature matrix? 2) If features are not given, then what is the meaning of train-test split in GNN? How do we predict on the test dataset if there are no features? If you can wrap your head around the above two questions, you will understand a great deal about the training process of GNNs and why it is very powerful compared to CNNs. In CNN, the training data and testing data are connected via features - the pixel values. The values of these pixels come from a Gaussian distribution. In GNN, even if features are not present for the nodes, the edges help us predict test data. The edges for the connection between training and test data. The data need not come from a Gaussian distribution always. We can still perform training and testing. This understanding will occur when you look at an actual dataset and try to implement the GNN framework. If you want to understand more, please check out this great lecture. You will how to implement a GNN from scratch in 30 minutes. Hope you will enjoy it. https://lnkd.in/gFZfMwsi Also feel free to check out Aiswarya's work at Togo AI Labs

134

𝐒𝐩𝐚𝐜𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐰𝐞𝐞𝐤 𝐨𝐧 Hugging Face 🤗 👨🏼‍🏫 Math Olympiad Solver (NuminaMath 7B TIR) The demo for NuminaMath: a series of language models trained to solve math problems using tool-integrated reasoning (TIR). The 6.96 billion parameter model won the first progress prize of the AI Math Olympiad (AIMO), with a score of 29/50 on the public and private test sets. 𝐖𝐡𝐚𝐭 𝐜𝐚𝐧 𝐲𝐨𝐮 𝐮𝐬𝐞 𝐭𝐡𝐢𝐬 𝐟𝐨𝐫? Calculate finance management outcomes, choose a workout plan by calculating gains, get a general algorithm to answer a specific question. All of these use cases can be best if you inform details, but the language model does have knowledge like ‘how much protein for muscle growth by kilogram’ (it considered 0.56 - 0.84 grams per kilogram). See all on today's edition of the TechTok Newsletter

2

Learnings from Day 15/58, •In any dataset, dependent #variables are usually present in the last column while all the rest features lie in the first n-1 columns. The .iloc[ ] #function selects the (row, columns). •Usage of the #scikitlearn #pythonlibrary by creating a sklearn #class and then using an #instance of the class ie #imputer to take care of the missing values. The #fit method of this class shall calculate the arithmetic value(exclude the string columns) while the #transform #method shall perform the replacement. •To use #onehotencoding to convert #categoricalvariables into binary numbers and hence also avoid any numerical order between the variables. --- Follow Sidhant Bhat! #xdayschallenge

12

🚀 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

How LLMs Work in 3 Sentences: 1) During training, LLMs digest sequential inputs of text data, which are converted into numerical representations called embeddings. 2) These embeddings capture the semantic meaning and contextual relationships between words, facilitating the model's understanding of language. 3) Through iterative adjustments of internal parameters via back-propagation, LLMs fine-tune their ability to predict the next word or sequence of words in a given context. Just as a chef combines ingredients like a bun, a burger, and cheese to make a cheeseburger, a large language model turns your words into embeddings, mixing them to generate useful responses.

1

1 Comment

Hey folks sharing a game-changing paper, "Attention is All You Need," which introduced the Transformer model. This work transformed NLP by using self-attention mechanisms to process sequences efficiently, eliminating the need for recurrence and convolutions. Key innovations: 🔍 Self-Attention: Captures dependencies across the sequence. 🔄 Multi-Head Attention: Focuses on different parts of the sequence simultaneously. 📐 Positional Encoding: Adds order information to tokens. This paper laid the groundwork for models like BERT and GPT. A must-read for anyone in AI! #NLP #MachineLearning #AI #Transformers

13

Here's my blog on the evolution of attention mechanisms. From Bahdanau's additive attention (2014) to Luong's multiplicative attention (2015) and Vaswani's self-attention in Transformers (2017). Blog:- https://lnkd.in/guWW5Eyb

31

2 Comments

📖 Check out this blog post by Nikita Agarwala, who walks you through the benefits of #BentoML with a practical example using the MovieLens dataset. She demonstrates how to build and serve a movie rating prediction model using simple #LinearRegression and BentoML. It is a typical example of how straightforward it is to use BentoML for serving 🤖️ models and creating REST endpoints. Link 👉 https://lnkd.in/gPVMh3XY

14

1 Comment

📢 Learn how AI revolutionizes data extraction, saving you time and increasing accuracy. 🛠️ #AIExtractsP&IDTags Read our latest blog on "Enhancing Data Extraction from Technical Drawings with AI" now! https://hubs.li/Q02GDm_N0

7

الحمدالله رب العالمين Project Name: Landmark Classification & Tagging for Social Media The "Landmark Classification & Tagging for Social Media" project aims to develop a system capable of automatically identifying and tagging landmarks in images uploaded to social media platforms. Leveraging deep learning techniques, specifically transfer learning with pre-trained models from computer vision libraries like Py Torch and torch vision, the project seeks to classify images into categories corresponding to various landmarks worldwide. By integrating advanced image processing and machine learning algorithms, the system will enable users to effortlessly tag their photos with accurate landmark information, enhancing the searchability and contextual relevance of images shared on social media platforms. This project not only addresses the challenge of automating image classification but also improves user experience by providing more informative and organized content on social media feeds. https://lnkd.in/dbTe83Eg #AWS #Udacity #machinelearning #deeplearning #projects #cnn #companies #it #informationtechnology #ai

7

May we all get the superpower to work with rich datasets📊 PS: Our team wrote a blog post on ‘How to build ML models?’ Do give it a read. Link here: https://lnkd.in/gF6nZ3gM

13

If we’ve a rich dataset, da da da da da, da da da da da da!  See all the insights we'd bring, no analysis could ever be boring. 'Cause we'd have all the data in the world! We’d have power to analyze, predict, and innovate, fueled by the wealth of information at our fingertips.  From market trends to customer behavior, every data point holds the potential for transformation. Check out our datakulture team’s latest blog post by Apoorva Grover on 'How to build ML models?' ⬇️ #Data #MachineLearning #BlogPost #ML #MLModel #DataLake #CloudApplications #sensor #IoT #RDBMS #MongoDB

11