Shahpar Yakhchi

I've picked up many advanced R tips and tricks from creating R packages which I regularly use for data analysis too 🎩🐇 This month I've distilled the most useful and fun tricks into a 3-hour tutorial that I'll be giving at WOMBAT in Melbourne on October 22nd. Learn how to {{embrace}} R's quirks to supercharge your analysis, avoid common pitfalls, and design wonderful R packages. Topics include: 👶 Basics: Revisit the basics of R and gain a deeper understanding of everything works 🎰 Vectorisation: Safely use vectors to calculate useful statistics, understand recycling and type casting 🪄 Metaprogramming: Using code to write, explore, modify, and evaluate code (how tidyverse design works!) 🌏 Melbourne, Australia (WOMBAT 2024 conference) 🕰 Tuesday 22nd October, 13:30 - 17:00 AEDT 💸 AU$150 (AU$75 for students) 🔗 Workshop information: https://lnkd.in/gGtY8pmq 🎓 Register here: https://lnkd.in/gNXrNwkA #RStats #DataScience #AdvancedR #opensource

57

4 Comments

🌟 Retrieval-Augmented Generation (RAG) is truly a game-changer in the AI landscape! By seamlessly combining retrieval-based and generative models, RAG architecture enhances both the accuracy and relevance of AI outputs. The detailed process involving chunking, embedding, and vector search ensures that responses are not only precise but also contextually rich. This innovative approach is particularly beneficial in providing enhanced relevance, improved accuracy, and versatility across various domains. RAG is setting a new standard for intelligent systems, making AI interactions more informed and reliable. #AI #MachineLearning #RAG #ArtificialIntelligence #TechInnovation #DataScience #VectorSearch #LLM #Embeddings #GenAI 🚀🤖📚

2

Best Practices in Retrieval-Augmented Generation (Fudan University, July 2024) Paper: https://lnkd.in/geCX9ZTB Abstract: "Retrieval-augmented generation (RAG) techniques have proven to be effective in integrating up-to-date information, mitigating hallucinations, and enhancing response quality, particularly in specialized domains. While many RAG approaches have been proposed to enhance large language models through query-dependent retrievals, these approaches still suffer from their complex implementation and prolonged response times. Typically, a RAG workflow involves multiple processing steps, each of which can be executed in various ways. Here, we investigate existing RAG approaches and their potential combinations to identify optimal RAG practices. Through extensive experiments, we suggest several strategies for deploying RAG that balance both performance and efficiency. Moreover, we demonstrate that multimodal retrieval techniques can significantly enhance question-answering capabilities about visual inputs and accelerate the generation of multimodal content using a "retrieval as generation" strategy."

21

If you are curious about Bayesian statistics and how/what it adds to traditional frequentist statistics or Machine Learning - check out this cool Aussie dude. Best explanations and guidance I have seen in a long time! Here is one example of logistic regression very well explained -

8

A new research paper from Apple shows that large language models (LLMs) struggle with basic math problems when irrelevant details are introduced, leading to significant drops in accuracy. The paper demonstrates that adding simple distractions, like a comment about the size of kiwis in a math problem, can confuse AI models, highlighting their inability to perform genuine logical reasoning. Researchers argue that AI models don’t reason like humans but rather replicate observed patterns from their training data, which leads to failures when problems deviate slightly from the norm. Some suggest that better prompts could improve performance, but the need for more context as distractions grow suggests limitations in AI’s current reasoning capabilities. The study sparks further debate on whether LLMs truly "reason" or just mimic the data they have been exposed to, raising concerns about how AI is being marketed and used today.

2

1 Comment

Simple and clear! Cyclical encoding to represent cyclical data (hours, month, moon cycles) through the use of trigonometric functions. https://lnkd.in/gqTiRK3s

1

Week 4 of DataTalksClub LLM-zoomcamp was rich in content. This week the focus was: 1) Setting up an LLM powered Streamlit app. The app provides users the capability to ask questions. We capture feedback from the user as part of the monitoring workflow. All data is stored and in a Postgres database. A visualisation dashboard is used for monitoring model performance (Grafana Labs). 2) We also explored how to use cosine similarity to deliver a verdict on how well an LLM is responding to user queries. A large language model was then used as a secondary evaluation method to guage quality of our user facing model. As always a big thanks to Alexey Grigorev and the team for making these available to many of us who enjoy collaborative learning and hands on projects. https://lnkd.in/drN-v7xv

12

Day 21: Reinforcement Learning: Shaping AI Through Rewards for #100DaysOfMLChallenge Today, we shift gears to explore reinforcement learning (RL), a fascinating area of AI inspired by how we learn through trial and error. Imagine an agent navigating a maze. In RL, the agent interacts with the environment, taking actions and receiving rewards (positive for good actions, negative for bad ones). Through this feedback loop, the agent learns to optimize its behavior to maximize the rewards. Here's what makes RL unique: No explicit instructions: Unlike supervised learning, RL agents don't receive step-by-step instructions. They learn through exploration and experimentation. Delayed rewards: The consequences of an action might not be immediate, requiring the agent to consider long-term goals. Adapting to dynamic environments: The environment can change, forcing the agent to continuously learn and adapt its strategies. The potential applications of RL are vast, with real-world examples already emerging in: Robotics: Training robots to perform complex tasks in dynamic environments. Game playing: AlphaGo, anyone? RL agents are mastering complex games like chess and Go. Resource management: Optimizing resource allocation in complex systems like logistics or energy grids. What are your thoughts on reinforcement learning? Are there any specific applications that excite you? Share your thoughts in the comments! #ReinforcementLearning #AI #MachineLearning #TrialAndError #Robotics

7

🌟 What an exciting exploration into accelerating text generation for large language models! 🚀 Your research is paving the way for faster and more responsive AI systems, opening up new possibilities for real-time applications. Speculative Decoding, Early Exiting Mechanisms, and Non-Autoregressive Methods are fascinating techniques that promise significant speedups in text generation while maintaining quality. It's inspiring to see collaboration between leading institutions like MIT and Columbia University driving innovation in AI efficiency. This paper is a valuable resource for researchers and practitioners alike, shaping the future of AI technology. Can't wait to delve into the full details and witness the impact of these cutting-edge techniques firsthand! 🧠✨ #AIResearch #TextGeneration #Innovation #MIT #ColumbiaUniversity #TechAdvancement 📚🔬

9

3 Comments

I recently worked on a project where I used innovation to categorize different banks in Iran's system banks. The results were astounding. Here's what I did. What innovative methods have you used in your projects? I’d love to hear your thoughts in the comments! 👇 #Innovation #DataScience #Banking #MachineLearning #Finance #Analytics #Iran #BusinessStrategy

3

learned about MMR today chroma has it, but solr I blv I'd have to write a custom function it's something I've been concerned about. cosimilarity in cosine embeddings and how to manage it. In ML, we would call it collinearity, but there isn't an easy way to derive a correlation matrix from a response to every other response in an embeddings query.

1

Informative demo video building ‘Automated Research and Summarisation agent’ using local LLMs(qwen-2.5, llama-3.2) through Ollama and LangGraph https://lnkd.in/gQR5z-hi #LearnAIML #LangChain #aiAgents #ollama #langGraph

6

📊 Categorical data is information divided into groups or categories. Encoding these variables is essential for model compatibility, mathematical operations, and enhancing model performance. Proper encoding improves accuracy and helps assess feature impact on predictions. 📺 Check out my latest video where I cover various categorical variable encoding techniques, including Label Encoding, Ordinal Encoding, One-Hot Encoding, Dummy Encoding, Binary Encoding, Base-N Encoding, and more. 🔍 Additionally, explore advanced strategies like Target Encoding to capture intricate feature-target relationships and learn how LightGBM optimizes performance by handling categorical variables intelligently. Video Link: youtu.be/MKuAQv6ybc8   Channel Link: youtube.com/@datatrek #datatrek #datascience #machinelearning #statistics #deeplearning #ai

3

As a continuation of my learning, I have completed a Intro to Deep Learning with Pytorch course on DataCamp

17

Exciting developments! 🐍 ML-Mamba’s linear complexity and efficient handling of long sequences could be a game-changer for multimodal tasks. The fact that it achieves comparable performance to top models like TinyLaVA and LLaVA, while using fewer parameters and being faster, is impressive. Looking forward to diving into the details and seeing how this innovation shapes the future of MLLMs! #MLMamba #AI #MultimodalAI #Transformers #MachineLearning #TechInnovation #LLMWatch #ArtificialIntelligence

1

Does it make sense to give LLM model context when making an inquiry? Yes, if the base model doesn't have enough background on subject of the inquiry. This isn't a novel idea, but it was a fun project to explore :) Built a RAG architecture (used one of Ollama models running on local system without web access). ChatGPT was used for knowledge base development & response evaluation. ChatGPT rated one query response as 7/10 for model with context, and 4/10 for model without context. This was not an exhaustive evaluation, just a few queries. This does mean, models' accuracies are now dependent upon quality of context. Which becomes easily visible when you run multiple queries through both models. What'll you find is, if your query doesn't have right keywords, the retrieval part of system might end up finding context that's not helpful or even causes issues for response generation, leading to entire model failing... Well this is where I am at. Project status: 🛠️ Merry Christmas 🎅

4

This is another powerful tool for AI engineers, helping businesses automate an increasing number of tasks. Here are the top 5 use cases where it can make the most impact: 1. Automated Document Processing: Efficiently extract data from contracts, loan applications, and financial reports to streamline workflows and reduce manual input in banking and financial institutions. 2. Invoice and Receipt Automation: Automatically read and categorize invoices, receipts, and expense reports, making financial reconciliation faster and more accurate. 3. Compliance and Regulatory Reporting: Extract key data from regulatory documents, ensuring accurate reporting for compliance purposes in industries such as banking and insurance. 4. Customer Onboarding: Automatically process identification documents, like passports or driver’s licenses, to speed up customer verification processes in sectors such as banking, telecom, and insurance. 5. Data Extraction from Scanned PDFs and Handwritten Notes: Handle complex documents like scanned contracts, handwritten meeting notes, or archival financial records, and convert them into structured, actionable data.

3

Absolutely spot-on analysis! 🌟 The next 6-8 months will indeed be pivotal for the AI field. What we’re about to witness is the culmination of an unprecedented scale-up in compute power, with new models trained on 5x-10x the resources compared to GPT-4-class models. It’s crucial to understand that the iterations we've seen lately—like GPT-4-Turbo, GPT-4o, GPT-4-Mini, Claude 3.5, and Gemini 1.5—are essentially refinements and optimizations of the same architecture that’s been around for 18 months. They’ve squeezed a remarkable amount of performance from post-training, but they’re still working within the same foundational limits. The real game-changer here is the emergence of vast GPU clusters. Building these clusters is no small feat—it involves overcoming enormous logistical challenges, from securing planning permissions and power supply to setting up advanced cooling systems and handling the intricate networking required to link tens of thousands of GPUs. This is why there’s a lag between the development of the base models and the next major breakthroughs. Once these new clusters are fully operational, we’ll see a true leap in model capabilities. GPT-5 is likely to be the first of these next-gen models, possibly arriving between November and February, followed by Claude-4, Grok-3, and Llama-4. But as you mentioned, we shouldn’t rush to judgment on how effective LLM scaling is based on the models we have now—they haven’t yet tapped into the full potential that these new clusters will unlock. In terms of capability, while it’s tempting to speculate, we’re in uncharted territory here. The advancements could be profound, but we’ll have to wait and see what these new models can really do. One thing’s for sure—the AI landscape is about to change dramatically, and we’re on the brink of witnessing something truly transformative. 🚀 Thanks for sharing these insights! 🙌 #AIRevolution #NextGenAI #LLMScaling #FutureOfAI #GameChanger

1

As part of MLx Rep. Learning & Gen. AI track of OxML. I had the pleasure of attending an enlightening lecture by Erin Grant from UCL titled "𝐍𝐞𝐮𝐫𝐚𝐥/𝐑𝐞𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧𝐚𝐥/𝐁𝐞𝐡𝐚𝐯𝐢𝐨𝐫𝐚𝐥 𝐂𝐨𝐦𝐩𝐚𝐫𝐢𝐬𝐨𝐧𝐬 𝐁𝐞𝐭𝐰𝐞𝐞𝐧 𝐇𝐮𝐦𝐚𝐧𝐬 & 𝐌𝐚𝐜𝐡𝐢𝐧𝐞𝐬." Erin shared her extensive research on comparing the behavior of vision AI models, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), with human vision. The lecture delves into the interpretability of these models relative to the human brain. While CNNs were designed with loose inspiration from the primate visual system and exhibit a structural similarity to the human brain, they often rely more on texture rather than shape for classification. In contrast, ViTs, though structurally different from the brain, tend to classify objects based on shape, which aligns more closely with human visual processing. Erin also discussed how ViTs demonstrate greater error consistency with humans and a higher shape bias compared to CNNs. This means that ViTs not only achieve higher accuracy but also make errors in ways that are more similar to human mistakes. She utilized metrics, such as Jensen-Shannon (JS) distance, to measure such similarities. Many thanks to OxML and AI for Global Goals for hosting this insightful module. #OxML24 #OxML #AI4GG #GlobalGoalsAI #RepresentationLearning #GenAI #AIResearch #Neuroscience #CNN #ViTs

20

3 Comments