Haitong Tao

Excited to share our latest work for time series clustering published in DMKD journal! Our model has linear time complexity w.r.t the number of instances and the time series length, and achieves SOTA performance across various time series data types. Check it out here: https://lnkd.in/ekmdcVbj #TimeSeries #AI #DeepLearning #MachineLearning #Research

6

LLMs can output tokens 200% faster using multi-token prediction. And it’s now been open sourced by Meta. Here’s how it works — Multi-token prediction extends standard language modeling -- it predicts several future tokens at once instead of just the next one. - First, the model uses a shared core to make a context representation from its past tokens. - Then, it predicts multiple future tokens in parallel using separate prediction layers. This approach is much more memory-efficient, so during training, it can handle large vocabulary sizes without overloading the GPU. And at inference, it can decode much faster. Why does it work so well? According to the paper, it’s because multi-token prediction is able to reduce the mismatch between training and inference. This makes the model better at handling long-term dependencies. Paper: https://lnkd.in/ge_quFaS HuggingFace models: https://lnkd.in/gUgzsbtZ #AI #LLMs

7

Dear network episode 8 is now live! 🚀 Explore the 4 stages of inference, cross-lingual sentiment analysis, the dynamics of auto bidding systems, cross-channel bidding, sequence recommendations modelling and a participatory approach to AI

8

Hi, Granger causality is an econometric hypothetical test for verifying the usage of one variable in forecasting another in multivariate time series data with a particular lag. However, it’s important to note that Granger causality doesn’t prove direct causation and should be interpreted cautiously when evaluating causal relationships between variables. For more details please read my blog on Granger causality explained using chicken and egg problem https://lnkd.in/gcKwkzph

3

Integrating Custom Data with an AI Agent Using LangChain 🦜 After some time and effort, I found an easy and effective way to integrate custom data with a Large Language Model (LLM). Here’s a short guide on how you can use your data to build an AI Agent using LangChain:

8

1 Comment

I’m excited to announce that I am starting a newsletter, 𝐓𝐡𝐞 𝐌𝐋 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬 (https://lnkd.in/gg4RV8tk). Having received many questions about how to succeed as a machine learning engineer at Big Tech companies, I realized the importance of mentorship in my growth. Through this newsletter, I aim to share my experiences, lessons learned, and offer mentorship to others. My goal is to provide actionable advice for both aspiring and experienced machine learning engineers. While I don't claim to have all the answers, I believe in the power of continuous learning and knowledge sharing. I’m committed to making this newsletter a valuable resource for anyone interested in machine learning. Your support and feedback will be invaluable. Subscribe now to read my first insight on 𝐇𝐨𝐰 𝐭𝐨 𝐒𝐮𝐜𝐜𝐞𝐞𝐝 𝐚𝐬 𝐚 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫: https://lnkd.in/gm_jcFXn. Sincere thanks to Raviraj Achar (https://lnkd.in/gq9a29ZG), Akash Mukherjee (https://lnkd.in/gCGTBRX2) and Kushal Shah(https://lnkd.in/gwArp3YC) for providing expert guidance on starting this newsletter. Check out their newsletters to learn more about leadership and career growth. #machinelearningengineer #careerinsights #ai #machinelearning

107

15 Comments

🚀 I'm happy to announce that our paper "Don't Just Translate, Summarize Too: Cross-Lingual Product Title Generation in E-commerce" has been accepted at the LREC-COLING 2024 ECNLP. Thanks the co-authors Taichi Nakatani, Stephan Walter, Daniel Vidal Hussey, Li Ling Tan for the collaboration! 🙂 Our study explores a joint translation and summarization task for cross-lingual product title generation in e-commerce. We introduce a novel data filtering technique to train a Machine Translation Model (<1B params) and compare it to an LLM Mixtral (7x22B params). We validated our results using Mixtral and found the automated evaluations aligned with human preferences. This work not only studies the capabilities of the open-source LLM Mixtral, but also demonstrates the potential for specialized, smaller models to outperform large language models in targeted tasks like cross-lingual product title generation. These findings highlight promising opportunities for developing efficient and effective NLP solutions to address specific business needs in the e-commerce domain. #amazonscience #LLM #machinetranslation #searchengine #ecommerce #translation #GenerativeAI #search #MultilingualSearch #mixtral #localization

125

7 Comments

Llama 3 adoption is scaling up quickly with lessons and cookbooks all over. A lot of versions of the mother model are coming out on HF. Fine tuning via approaches such as QLoRA on your own data ie turning the generalist into a domain specific expert is a cheap option over time.

5

Sharing our new paper-“Blended RAG: Improving RAG (Retriever-Augmented Generation) Accuracy with Semantic Search and Hybrid Query-Based Retrievers” I'm thrilled to announce the release of my first paper! We've been working hard to improve question-answering systems, especially when dealing with sensitive business data. Instead of just focusing on one part of the system, we made improvements to both how it finds relevant information and how it generates answers. Our new method, called "Blended RAG," combines different techniques to search for information more effectively. We tested it on various datasets, including NQ and TREC-COVID for information retrieval and SQUAD for question answering. The results were impressive, with our method achieving 87% accuracy on NQ and TREC-COVID datasets and 57% on SQUAD, setting new benchmarks in the field. This scalable method has proven effective across millions of documents and is leading the charge in RAG methodologies. A big shoutout to my co-authors Shivam Solanki and Kunal Sawarkar for their incredible work! Check out our findings and support our work: View Benchmarks: https://lnkd.in/g2Bcpk5r Read the Paper: https://lnkd.in/gEWUYM83 Star Our Repo: https://lnkd.in/g4_euEE2 #llm #rag #IR #generativeai #ai

37

🤖 Unleash the Power of LLMs with LangChain: Building a T-Shirt Database Q&A Bot 👕 In today's data-driven world, the ability to extract insights from databases is crucial, but interacting with databases often requires a deep understanding of SQL syntax and database schemas. What if you could simply ask natural language questions and get accurate answers from your database? That's exactly what I've explored in my latest Medium article, where I walk through building a T-shirt database Q&A bot using LangChain, a powerful library for creating applications with large language models (LLMs). In this article, you'll learn how to: ✅ Set up an LLM integration with a SQL database   ✅ Create a user-friendly interface with Streamlit   ✅ Leverage few-shot learning and vector embeddings   ✅ Define prompts and templates for LLM interaction   ✅ Implement the SQLDatabaseChain for natural language queries  Whether you're a T-shirt retailer, a data analyst, or a developer, this approach has the potential to transform the way you interact with databases, empowering you to extract insights and make data-driven decisions with greater ease and accuracy. Interested in learning more? Check out the full article here: https://lnkd.in/gVhH9F8U Special thanks to Codebasics for the tutorial that guided me through the learning process for this project. Let me know your thoughts and experiences with LLMs and database interaction in the comments! #LLM #LangChain #NaturalLanguageProcessing #DatabaseManagement #DataAnalytics #Codebasics

16

2 Comments

🚀 Optimizing Large Language Models for Enhanced Performance and Efficiency In the rapidly evolving world of AI, the optimization of Large Language Models (LLMs) stands as a crucial frontier for achieving both high performance and operational efficiency. Today, I’m excited to share a comprehensive guide from Humanloop, detailing the latest strategies in optimizing LLMs. 📊✨ Key highlights from the blog: Efficient Training Techniques - Learn about the innovative methods to train LLMs more effectively, reducing computational overhead and costs without compromising on quality. Fine-Tuning for Specific Tasks - Discover how targeted fine-tuning can drastically improve model performance on specific tasks, making them more relevant and useful in practical scenarios. Quantization and Pruning - Explore how these techniques help in reducing the model size and increase inference speed, making deployment easier and more cost-effective. Utilizing Hardware Accelerators - Understand how leveraging the right hardware can lead to significant performance boosts and efficiency gains. This guide is not just theoretical—it provides actionable insights that can be applied right now to harness the true potential of LLMs in various applications from chatbots to sophisticated data analysis tools. For anyone involved in AI and machine learning, especially those working with large language models, this is a must-read! Dive into the full article for a deeper understanding and practical tips: https://lnkd.in/dv5mfVW8 #AI #MachineLearning #LanguageModels #TechnologyInnovation #DataScience #LLM

10

Do you want to build and deploy custom LLM agents from scratch? Check out my blog post with Databricks Mosaic Research to read about how JetBlue is shipping robust GenAI solutions with DSPy and Databricks! https://lnkd.in/gKcVKQ_B

82

4 Comments

𝐏𝐲𝐭𝐡𝐨𝐧 𝐅𝐫𝐢𝐝𝐚𝐲! 𝐓𝐨𝐝𝐚𝐲, 𝐰𝐞’𝐥𝐥 𝐥𝐞𝐚𝐫𝐧 𝐡𝐨𝐰 𝐭𝐨 𝐮𝐬𝐞 𝐆𝐫𝐨𝐮𝐩 𝐁𝐲 𝐢𝐧 𝐨𝐮𝐫 𝐏𝐚𝐧𝐝𝐚𝐬 𝐚𝐠𝐠𝐫𝐞𝐠𝐚𝐭𝐢𝐨𝐧𝐬! Welcome to Lesson 10 of #SQL_to_Pandas and Day 6 of Aggregations! For those just joining us, review what we've learned about aggregations so far: ⤷ Lesson 5 - Aggregating with SUM: https://lnkd.in/gh3fnKGQ ⤷ Lesson 6 - Aggregating with Mean, Min, Max: https://lnkd.in/gXTtAJBf ⤷ Lesson 7 - Count Aggregation: https://lnkd.in/gNPgdvZ6 ⤷ Lesson 8 - Multi Aggregations on Single Column: https://lnkd.in/gT6TZTmf ⤷ Lesson 9 - Multi Aggregations on Multiple Column: https://lnkd.in/gJbUdqrX Today, we’ll explore how to group data using the Group By method. In SQL, you use the GROUP BY clause to group data by one or more columns and then apply aggregation functions to these groups. This helps in summarizing data based on categories. In Pandas, we achieve the same results using the .groupby() method. Imagine you have a DataFrame with company data and want to group it by industry to calculate the average number of employees in each industry. Grouping data helps you gain insights into patterns and distributions within your dataset, enabling you to make data-driven decisions more effectively. 👋 Follow me, Shane Butler, to subscribe to the #SQL_to_Pandas series! 💬 Comment '#SQL_to_Pandas' or ♻ Repost if you’d like me to send you a PDF of all the material covered so far. #SQL_to_Pandas Lesson 11 on Inner Joins up next. 🚀 #datascience #pandas #python #sql #dataanalyst #businessanalyst #careergrowth #mentorship #sqltopandas

61

4 Comments

Top 3 Learnings from Chip Huyen O'reilly podcast. 🌟 2 reasons limiting most of the companies from using Gen AI in Customer facing applications: Reproducibility, Compliance (GDPR,EU AI ACT etc) 🌟 2 ways most companies approach Gen AI: Problem first: We have this problem , Can we use Gen AI to solve it better . Chaos Monkey Research: Ask all your Senior to start using Gen AI foundation models, test and come up with new use cases But having Applied Research team for LLM's doesn't make it sense 🌟AI Engineering : Because we have good foundation models and spend lot less time on model development AI engineers are expected to be full stack engineers (end to end) with knowledge of front end engineering etc https://lnkd.in/ez7TEQAC

4

I am proud to share my latest article “330 Weeks of Data Visualizations" 🎉 This journey has been a labor of love and dedication, starting back in 2018 when I first dived into the world of data visualization. Week after week, I challenged myself to create a new visualization, explore diverse datasets, and hone my skills. Now, over 330 weeks later, I’m excited to reflect on this journey and share the key takeaways that have shaped my career and personal growth. If you’re into data visualization or looking to boost your data science skills, I hope my story inspires you to start your own journey. Thank you Towards Data Science for publishing another article from me! #DataVisualization #DataScience #Tableau #CareerDevelopment

66

1 Comment

Learning and using LangChain as an LLM framework is truly helpful in understanding and running LLM into the production phase and evaluate the model.

3

Sharing my latest article on the RAFT: Teach LLMs to be better at RAG Discover how RAFT bridges the gap between fine-tuning and RAG, enhancing model performance by integrating domain knowledge with powerful retrieval techniques. 🔍 Dive into: • The mechanics of RAFT • Benefits over traditional methods • Real-world applications using Meta Llama 2 Read the full article here: https://lnkd.in/g7ZQNE5s #AI #MachineLearning #DataScience #NaturalLanguageProcessing #AIInnovation #MetaLlama2 #RAFT #llm #RAG #genAI #LargeLanguageModels

9