Mustafa Mudassir’s Post

Yesterday the National Academy of Sciences held their "AI Day for Federal Statistics: CNSTAT Public Event in DC. #MindBlown Key Takeaways (for me): ✅ DOMINANT USE CASES There was a clear and dominant focus on #DataQuality!!!! > Classifying and coding (cleaning up dirty data) > Document processing (data extraction) > Record matching (data linkage for the purpose of data cleansing) > Data imputing (predicting missing data) > Meta data enrichment (structure cleaning, standardizing, etc.) > RAG queries to develop and test for more accurate GPT responses ✅ REGS, POLICIES, LEGISLATION, ETC. Legal parameters were mentioned in EVERY talk and session. > Information Quality Act > Foundations of Evidence Based Policy Act of 2018 > EO 14110: Safe, Secure, and Trustworthy Development and Use of AI > EO 13859: Maintaining AI leadership > M-24-10: AI Governance for Agencies > M-19-23 & M-21-27: Evidence-based evaluation, learning agendas > M-19-15: Information Quality > M-14-03: Administrative Data Security > M-19-18: Federal Data Strategy > A-130: Managing Information ✅ RESOURCES It's all about knowing where to find resources!! > Data.gov (293,975 datasets available) > NIST Collaborative Research Cycle Data & Tools: https://lnkd.in/dK8DTRwB ✅ BEST QUOTE And probably one of the best quotes I've heard to explain AI came from Ken Cunningham of Microsoft Federal when he said: #AI is to #Data + #Math as #Humans are to #Carbon + #Water. #NailedIt Event Link: https://lnkd.in/d2mNMmH7 {in case they offer a report out or recording of the sessions}

🚀 With the rise of GenAI, the concept of synthetic data has gained even more significance… 📊 As companies engage in the new arms race to develop LLMs with parameters in the billions, their hunger for data intensifies. In this context, synthetic data plays a crucial role. Already in 2021, Gartner predicted that by 2030, most of the data used in AI will be artificially generated by rules, statistical models, simulations, or other techniques. 💡 That's why I had to revisit this topic and its dual use in the context of GenAI which indeed is capable of generating multi-modal content and data but also relies heavily on synthesized data during its training. 🔗 If you want to learn more about the fascinating concept put also potential vicious cycle this bears, click here: https://lnkd.in/eAQ6wHdi #AI #ArtificialIntelligence #GenAI #LLMs #Data #BigData #Analytics #DataAnalytics #MachineLearning #ML

Organizations are increasingly turning to synthetic data for research, testing new developments, and ML research purposes. Synthetic data enables researchers to create large datasets without the constraints of real-world data, fostering extensive machine learning (ML) training and testing. Learn more about this innovative approach in the paper: https://lnkd.in/gQe46k2T https://lnkd.in/gBZRSPeu.

We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’ We demonstrate that it must be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web. Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet. excerpt from AI models collapse when trained on recursively generated data, Nature, 2024-July-24, Springer Nature Group https://lnkd.in/gGXH9kw8 ------ We're running out of people generated data to train the models. Synthetic data provides an alternative with many positives. But we have to beware the AI model collapse over time based on accumulation and amplification of mistakes. Buckle up. Curvy roads ahead. #data #ai #genai #llms #llm #train #peopledata #learn Maribel Bill #syntheticdata #limits #nature

Particularly with numerical data or time-sensitive facts LLMs often produce factually incorrect information due to their probabilistic nature and gaps in their training data. Advancements in retrieval mechanisms like Retrieval-Augmented Generation (RAG) have improved LLMs by fetching external data before generating responses, but RAG's static approach limits its ability to handle complex queries. Retrieval-Integrated Generation (RIG) overcomes this by dynamically retrieving data during response generation, refining answers in real time. The shift to Retrieval-Integrated Generation (RIG) addresses the need for AI systems capable of real-time problem-solving, making it more adaptable to evolving queries compared to traditional models like RAG. RIG’s iterative retrieval process is well-suited for dynamic applications, such as finance, where real-time insights are crucial. However, RIG's continuous data retrieval demands more computational resources, potentially increasing latency and infrastructure costs, especially in large-scale deployments. Despite these challenges, its benefits in handling complex, real-time queries make it a valuable advancement. #ai #rig #google #llms #DataGemma #datacommons

🌐 From Retrieval to Intelligence: Agent+RAG and TruLens Are Revolutionizing AI Workflows In AI, retrieval systems like RAG are evolving to meet real-world challenges—solving multi-step reasoning, ensuring reliability, and adapting to dynamic workflows. Agent+RAG and TruLens are redefining what’s possible. What’s New? Agent+RAG enhances intelligence by: ✅ Enabling dynamic reasoning for multi-step tasks. ✅ Adapting in real-time with tools like Kafka and Databricks for scalable workflows. ✅ Powering innovations like fraud detection, supply chain forecasting, and personalized assistants. TruLens brings transparency with metrics for: 1️⃣ Groundedness: Are outputs factually accurate? 2️⃣ Context Relevance: Does retrieved data fit the task? 3️⃣ Accuracy: Can answers be verified against trusted sources? Why It Matters: Agent+RAG doesn’t just retrieve—it reasons and adapts, empowering developers and data engineers to build trustworthy, scalable systems for industries like healthcare, finance, and logistics. 📖 Dive deeper: From Retrieval to Intelligence: https://lnkd.in/gvafBcMG 💡 How do you see these tools transforming your workflows? Let’s discuss! #AI #AgentRAG #DataEngineering #Kafka #Databricks #Innovation

The iQuantNY blog was a great success illuminating what's actually in open data. Inspired by that, I had a little fun creating a quick gpt-enabled blog post tongue-in-cheek-titled aiQuantCA. It's a little analysis of groundwater levels from the data.ca.gov open data portal. It's... kinda interesting? But then not really useful per say. It certainly has the first pass appearance of looking smart. Sometime AI can be helpful with tasks, sometimes it feel like this sort of open ended inquiry isn't exactly increasing the signal to noise ratio. https://lnkd.in/giUktv4e Thoughts for future refinements: I could see AI particularly succeeding in more semi-structured tasks like monitoring a certain dataset and flagging outliers based on certain criteria. This type of open ended inquiry is getting shinier as gpts progress but then isn't really insightful per say.

🟢 Free PDF: "𝗖𝗵𝗲𝗮𝘁𝘀𝗵𝗲𝗲𝘁 𝗼𝗻 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮 𝗦𝘆𝘀𝘁𝗲𝗺𝗮𝘁𝗶𝗰 𝗔𝗜/𝗠𝗟 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲" ~ Pandey 1. Ideate: Define the problem & research solutions. 2. Manage Data: Identify data sources to address imbalances. 3. Develop: Choose algorithms, implement, train, and fine-tune your model. 4. Integrate: Design APIs and integrate AI with backend services. 5. Secure: Conduct threat modeling and implement robust security measures. 6. Deploy: Set up cloud infrastructure and implement CI/CD pipelines. 7. Monitor: Track performance, user interactions, and model drift. 8. Maintain: Review code regularly and maintain database. 9. Scale: Implement strategies to handle growing demands. 10. Assess Ethics & Compliance: Assess AI ethics, bias, and ensure regulatory compliance. #AI #datascience #machinelearning source: Brij kishore Pandey ________________________ 🟡 Follow Martin Roberts, PhD for other great resources that I've found on the internet 🔴 Visit my blog "𝗘𝘅𝘁𝗿𝗲𝗺𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴" https://lnkd.in/g5d965kh. Blog Topics include: * 𝗔 𝗻𝗲𝘄 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺 𝗳𝗼𝗿 𝗺𝘂𝗹𝘁𝗶-𝗮𝗿𝗺𝗲𝗱 𝗯𝗮𝗻𝗱𝗶𝘁𝘀 https://lnkd.in/gCTnWFvb * 𝗧𝗵𝗲 𝘂𝗻𝗿𝗲𝗮𝘀𝗼𝗻𝗮𝗯𝗹𝗲 𝗲𝗳𝗳𝗲𝗰𝘁𝗶𝘃𝗲𝗻𝗲𝘀𝘀 𝗼𝗳 𝗾𝘂𝗮𝘀𝗶𝗿𝗮𝗻𝗱𝗼𝗺 𝘀𝗲𝗾𝘂𝗲𝗻𝗰𝗲𝘀 https://lnkd.in/gfiG_Edx * 𝗛𝗼𝘄 𝘁𝗼 𝗲𝘃𝗲𝗻𝗹𝘆 𝗱𝗶𝘀𝘁𝗿𝗶𝗯𝘂𝘁𝗲 𝗽𝗼𝗶𝗻𝘁𝘀 𝗼𝗻 𝗮 𝘀𝗽𝗵𝗲𝗿𝗲 https://lnkd.in/gEu6HjxZ * 𝗛𝗼𝘄 𝘁𝗼 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝘂𝗻𝗶𝗳𝗼𝗿𝗺𝗹𝘆 𝗿𝗮𝗻𝗱𝗼𝗺 𝗽𝗼𝗶𝗻𝘁𝘀 𝗼𝗻 𝗵𝘆𝗽𝗲𝗿𝗱𝗶𝗺𝗲𝗻𝘀𝗶𝗼𝗻𝗮𝗹 𝘀𝗽𝗵𝗲𝗿𝗲 𝗮𝗻𝗱 𝗯𝗮𝗹𝗹𝘀 https://lnkd.in/gPmE5xMQ and many more! ;)

"Why can't I use the machine learning powering my #MultiTouchAttribution to power #MMM? It's a fair question. I'm going to attempt to hit a sweet spot between #TMI and not enough, so here it goes. 1. Machine Learning, like AI itself, covers a multitude of subsets of intelligence. But the ML that most people are most used to is all about #patternrecognition. And it's absolutely true that ML can find big and small patterns that would elude the human brain. The problem is that while we can engineer a pattern of behavior, that doesn't mean the behavior causes the desired effects. It just means that we succeeded in getting a lot of people to do the same thing. The Max Planck Institute for Intelligent Systems, Montreal Institute for Learning Algorithms (MILA), and Google Research delved deeply into the challenges related to causal representations in machine learning models, publishing a paper in 2021. "The majority of machine learning algorithms boil down to large scale pattern recognition." That's still the case. In short, ML inaccurately conflates patterns with what's important, causally speaking. 2. Machine learning can't cope with #TimeLag. It doesn't factor in the time that passes between the effects in a pattern and what caused those effects. Without time lag, you have nothing that enables you to optimize spend, particularly in B2B where the time lags are more significant. 3. Machine Learning doesn't deliver a multivariable view of what's happening. In #MTA or machine learning-based MMM (not really MMM), the pattern matching is about outcomes, not what caused those outcomes. It involves neither #correlation nor #regression, both mainstream approaches that move us toward better understanding of causal relationships. 4. #MachineLearning is very vulnerable to what are called "counterfactuals." An example is that you know that when a baseball bat hits a baseball, the player's arm powered the bat. But a ML algorithm can't see it or know it in and of itself. What this means is that the "cause and effect" often gets inverted in a machine learning model, with the bat being credited for making the arm swing. 5. ML doesn't capture the interdependency of factors. It assumes the observations are not interdependent and have a constant probability of occurring. That's called #iid in machine learning parlance. It's a problem. 6. ML operates on the same general premise as #GenAI: it predicts based on statistical regularities, i.e. "Past is Prologue." As soon as things get volatile, ML accuracy is upended as we saw with ML models during Covid. Many ML models began to fail because they had been trained on statistical consistencies instead of variabilities. This characteristic is also why GenAI does better with established, unchanging fact sets than with speculative ideas. More to come on this.

What is 𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥-𝐀𝐮𝐠𝐦𝐞𝐧𝐭𝐞𝐝 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧 (𝐑𝐀𝐆)? Imagine this: you’re asking the model something complex, and instead of just digging through what it learned months (or even years!) ago, it actually goes out, finds the freshest info, and brings it right back to you in its answer.  That’s Retrieval-Augmented Generation (RAG) in action. RAG is like an AI with a search engine built in.  Instead of winging it with just its trained data, it actively pulls in real-time facts from external sources and combines them with its own insights. The result? You get a response that’s not only coherent but packed with relevant, up-to-date information. How it works? 1. Query encoding: When a user inputs a question, it’s encoded into a format that a search engine or database can process.  The encoding turns the question into a vector or "embedding". 2. Retrieval phase: The retriever component then “searches” within an external database or document repository for relevant information. This step is critical as it brings in fresh, factual data, unlike traditional models that rely solely on pre-trained knowledge. The retrieved documents, often ranked by relevance, provide context for the response. 3. Generation phase: The embedding model takes both the initial query and the retrieved information. It compares these numeric values to vectors in a machine-readable index of an available knowledge base. Then it finds a match or multiple matches and retrieves the related data, converts it to words, and passes it back to the LLm. 4. Response generation: With retrieved data, LLM combines the retrieved words and crafts response as an answer to the user. Pros and Cons ➕ Pros: real-time access, improved accuracy, reduced hallucination, transparency ➖ Cons: complex implementation, increased latency, resource-intensive, dependency on data quality #ai #ml #llm #rag #techwithterezija