Sriram Raghavan

If you are still thinking that AI is just a hype, you should listen the conversation between Mukesh Ambani and NVIDIA CEO Jensen Huang. Why ? Mukesh Ambani bullishly talks about Generative AI models, GPUs, need of open source models etc. He takes up the challenge to surprise world with AI capabilities in coming years. He is not someone who has been leading IT/Software companies day in day out like Narayan Murthy/ Shiv Nadar/Bill Gates etc. What it means? AI has already arrived and every big firm, irrespective of their traditional business, wants to build business and create impact using this. AI is not just another cool tech anymore. It is the industrial revolution changing every facet of our life. Embrace it or get ready for classic "Kodak Case study" in your career and business. Watch here: https://lnkd.in/ejvnCMHZ

10

Researchers from Penn Engineering, IISc, and MIT have developed a groundbreaking low-energy method to induce amorphization in indium selenide wires, potentially revolutionizing phase-change memory (PCM) technology by drastically reducing power requirements for data storage applications. This discovery opens the door to new, energy-efficient memory devices for widespread use. #Science #Technology #Innovation #Engineering #BenChanTX How do you think advancements in low-power memory technology could impact data-heavy industries or everyday devices?

1

Looking forward to this panel where Divakar Tantravahi, Meduri Ravi Kumar, and Cosimo Spera will talk about the latest AI technologies that are changing product and customer support, along with their real world examples and practical strategies. If you have not registered, use this link - https://lnkd.in/gdTMgryA #ai #businessoutcomes #product #customer

22

Good video for non-experts in #ai to watch. I 100% agree with their points, and have thought similar things for over a year now on #genai and with #selfdriving for years. When I founded my company called Pragmatic AI Labs , in 2016 the focus was "pragmatic", i.e. realistic, skeptical and useful. 80% of GenAI or "AI (Marketing Term)" is none of these. Examples: A. Self-driving in current form is likely a bust, we need new roads for smart cars B. GenAI is a power drill being sold as an airplane....useful, but 95% less useful than told C. Much of what is called "AI", is just remarketed linear regression D. Consider the source...#bayarea #tech people lie; the end. E. Love the call out to use #criticalthinking https://lnkd.in/eUZAJzUG #ai #snakeoil

5

1 Comment

💥💥💥 $100K or 100 Days: Trade-offs when Pre-Training with Academic Resources Apoorv Khandelwal, Tian Yun, Nihal V. Nayak, Jack Merullo, Stephen H. Bach, Chen Sun, Ellie Pavlick Abstract Pre-training is notoriously compute-intensive and academic researchers are notoriously under-resourced. It is, therefore, commonly assumed that academics can't pre-train models. In this paper, we seek to clarify this assumption. We first survey academic researchers to learn about their available compute and then empirically measure the time to replicate models on such resources. We introduce a benchmark to measure the time to pre-train models on given GPUs and also identify ideal settings for maximizing training speed. We run our benchmark on a range of models and academic GPUs, spending 2,000 GPU-hours on our experiments. Our results reveal a brighter picture for academic pre-training: for example, although Pythia-1B was originally trained on 64 GPUs for 3 days, we find it is also possible to replicate this model (with the same hyper-parameters) in 3x fewer GPU-days: i.e. on 4 GPUs in 18 days. We conclude with a cost-benefit analysis to help clarify the trade-offs between price and pre-training time. We believe our benchmark will help academic researchers conduct experiments that require training larger models on more data. 👉 https://lnkd.in/d5Vh3ycg #machinelearning

1

Organizations have insufficient technical capabilities for the speed and scale of their ML model implementations. My colleague Sumit Agarwal just published a very helpful research with a capability view and sample vendors for defining a modular, flexible and scalable MLOps solution. https://lnkd.in/gnB_T9bq #MLOps #LLMOps #AI

149

17 Comments

Lifting the Veil on AI's decision making, Vinay Kumar Sankarapu from Arya.ai will share about AI transparency that goes beyond the usual "black box" discussions. In today’s world of complex deep learning models, understanding why a model makes certain decisions is crucial—especially in mission-critical applications. He'll explore DL-Backtrace, a new model-agnostic technique designed to bring transparency to any deep learning model, whether it’s an LLM, computer vision, NLP, or tabular data model. What You’ll Learn: - The importance of explainability for mission-critical deep learning applications - How DL-Backtrace compare to existing techniques like SHAP, LIME, and GradCAM for models including LLMs (Llama 3.2), NLP (Bert), and CV (ResNet) - Real-world benchmarking results showing DL-Backtrace’s performance across varied deep learning architectures - Insights into the current limitations and future possibilities in scaling explainability for large models #ResponsibleAI MLOps World Generative AI World

7

While Linkedin Space is overwhelmed with GenAI contents , there has been an excellent drive going on by Scientific community on how AI (Neural Net and other DL/ML s) can help solving science probems of solving equations and simulations. It opened my eyes on vast opportinuty of possibilities. For further, please read Physics Inspired Neural Net, Neural Operator ... More here .. https://lnkd.in/eNRtswmr (X page of Anima Anandkumar ) https://sciml.ai/ #sciml #aiforscience

18

1 Comment

Recommend this course to understand better when to use pre-training vs fine-tuning of LLM Models

11

Here is an interesting paper from our Data Science & Insights department (Megha Sinha, Sreekanth Menon, Ram Sagar), published recently at IEEE  https://lnkd.in/gZEKvzCd.  Drawn from practical experiences at Genpact you can use this as a guide for effectively managing and deploying LLMs, ensuring they meet the desired performance standards and remain reliable and trustworthy in their outputs. The paper lists the following steps in the LLMOps framework for an enterprise-grade Gen AI solution: 1. Defining a Project Charter: Create a well-defined project charter with clear objectives and expectations. 2. Generative AI Use Case Selection: Assess use case viability using a qualification matrix considering risk and ROI. 3. Solution Architecture: Collaborate with SMEs and ML engineers to develop robust architectures. 4. Design Data Pipeline: Develop a comprehensive data strategy and identify data sources. 5. Data Collection: Implement automated data collection scripts and normalize data formats. 6. Data Transformation: Correct inconsistencies and remove duplicate data using imputation techniques and a combination of human-annotated and synthetic data. 7. Data Storage: Choose between open-source and closed-source vector databases based on performance factors. 8. Data Handling: Use data versioning and encryption with access controls for secure data handling. 9. Model Training: Optimize model performance using established libraries and techniques. 10. Model Evaluation: Measure accuracy, coherence, contextuality, factuality, robustness, and trustworthiness. 11. Model Governance: Define metrics to capture at each run and set thresholds for re-training. 12. Model Inference: Manage issues like concurrent users, scalability, and performance for off-the-shelf LLM usage. 13. Model Serving: Consider costs and optimal use of compute resources. 14. Model Monitoring: Capture machine and human feedback to improve LLM performance.

16

💥💥💥 Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization Phillip Guo, Aaquib Syed, Abhay Sheshadri, Aidan Ewart, Gintare Karolina Dziugaite Abstract Methods for knowledge editing and unlearning in large language models seek to edit or remove undesirable knowledge or capabilities without compromising general language modeling performance. This work investigates how mechanistic interpretability -- which, in part, aims to identify model components (circuits) associated to specific interpretable mechanisms that make up a model capability -- can improve the precision and effectiveness of editing and unlearning. We find a stark difference in unlearning and edit robustness when training components localized by different methods. We highlight an important distinction between methods that localize components based primarily on preserving outputs, and those finding high level mechanisms with predictable intermediate states. In particular, localizing edits/unlearning to components associated with the lookup-table mechanism for factual recall 1) leads to more robust edits/unlearning across different input/output formats, and 2) resists attempts to relearn the unwanted information, while also reducing unintended side effects compared to baselines, on both a sports facts dataset and the CounterFact dataset across multiple models. We also find that certain localized edits disrupt the latent knowledge in the model more than any other baselines, making unlearning more robust to various attacks. 👉 https://lnkd.in/dpfrUDNr #machinelearning

1

GenAI leads to new economies of scale by generating artifacts in bulk or by helping humans to generate them faster. There is plenty of room for commercialized solutions to create new and original value with those #GenAI artifacts. 

96

3 Comments

Unleashing GenAI Innovation: Insights from the 3AI #SPECTRA Roundtable I had the pleasure of moderating the roundtable discussion on "RAG Strategies: Leveraging Diverse Techniques for GenAI Innovation" at the recent 3AI SPECTRA event. As the moderator, I posed these thought-provoking questions to our esteemed panelists: 1. How do enterprises prioritize their strategies to drive business growth and efficiency — focusing on increasing revenue or reducing costs through improved productivity? 2. what is the typical preference for optimizing processes — agentic workflow, RAG approach, or a combination of both? 3. What do enterprises value more when interacting with systems — achieving high-quality responses, low-latency per request, or finding a balance between to meet business expectations? 4. What is the prevailing preference for an enterprise — utilizing open-source models or investing in commercial models for implementing Gen AI solutions? 5. When enterprises kickstart Gen AI projects, what approach do they often take— starting with one Proof of Concept and then scaling it up, or experimenting with 2-3 POCs and then deciding which one to scale based on outcomes and feasibility? 6. What priorities do enterprises have when it comes to system performance —minimizing costs per request, delivering high-quality responses consistently, or achieving a blend of efficiency and effectiveness? 7. Which type of models, small domain-specific or large general models typically deliver better results and value for enterprises in terms of accuracy, scalability, and adaptability to business needs? Listen to the playback video for key takeaways and the insightful conversation with our esteemed panelists. A huge thank you to my co-moderator Shibaji Chatterjee and our esteemed panelists Ashok Kalra, Karan Makim, Kiranjit Pattnaik, Prashant Kumar Jha, and Shivani Sharma for sharing such valuable insights. My deepest gratitude to Sameer Dhanrajani, Nandakumar Ramaiah, and Kapil Gandhi for organizing this incredible event and providing me with this amazing opportunity. #GenAI #AI #3AISpectra #Innovation

16

A paper titled "Evaluating LLMs’ Effectiveness in Culturally Nuanced, Low-Resource Real-World Scenarios" by Millicent Ochieng, Varun Gumma, Sunayana Sitaram, Jindong Wang, Vishrav Chaudhary, Keshet Ronen, Kalika Bali, and Jacki O'Neill investigates the performance of seven Large Language Models (LLMs) in sentiment analysis using a dataset derived from WhatsApp chats in English, Swahili, and Sheng (a mix of English and Swahili). Key findings include: 1. Performance Variation: GPT-4 and GPT-4-Turbo excelled in understanding diverse linguistic inputs and context, showing high consistency and transparency, whereas other models like Mistral-7b and Mixtral-8x7b achieved high F1 scores but struggled with linguistic nuances and transparency. 2. Cultural Nuance Challenges: All LLMs had difficulties incorporating cultural nuances, especially in non-English settings, indicating a need for continuous improvement in handling low-resource languages and culturally diverse data. 3. Evaluation Methodology: The study combined quantitative analysis (F1 scores) with qualitative assessment of LLMs' explanations, highlighting the importance of qualitative Human-Computer Interaction (HCI) methods in understanding model performance in real-world, multilingual scenarios.

24

1 Comment

Advancing multilingual AI in Asia : the topic of my talk in the recently concluded NVIDIA AI summit in Mumbai really talks about how Tech Mahindra and NVIDIA together are taking LLMs for the new world order Yes there are multiple LLMs but as you all know foundation models do very well in a generic space but fail when its verticalized domain based space . They have to be fine tuned . Data has to be curated well and above all it’s not science but sort of a dark art to attain resonance of models which are guardrailed , and are ethical for an organisation. The future is just not verticalized LLms but genetic workflows as well , but the biggest area of work today is how to give reasoning and context to these models so that they establish a world model in their dreams to answer correctly. Stay tuned …. https://lnkd.in/dacH_r2Y

132

2 Comments

Dive into the world of #ecommerce innovation with the latest article by Sriram P H., CEO and Co-Founder, DaveAI, as we explore the integration of large language models (LLMs) with recommendation engines. Discover how this cutting-edge technology is revolutionizing customer experiences and driving sales in the digital marketplace. From personalized shopping assistants to predictive product recommendations, the possibilities are endless.   Read the full article exclusively on Express Computer to uncover the future of conversational commerce and personalized recommendations in e-commerce: https://lnkd.in/gNtNjzeh #ai #ecommerceinnovation #customerexperience

56

1 Comment

😎 Big challenge alert 🚀: How to disrupt the GPUs market! Researchers and tech enthusiasts are exploring innovative ways to simplify Attention and steer clear of matrix multiplication complexities. From classical hashing to ternary matrices and now clipping, strategies are evolving to transform multiplication into addition with impressive results. If you're passionate about deep tech or considering a startup, this is a prime problem to tackle. Dive into the details here: 🌟 #TechInnovation #GPUs #ResearchOpportunity

40

2 Comments

Developing #GenerativeAI models & #LLMs that reflect the linguistic & societal diversity of India requires extensive research and comprehensive data collection. This necessitates significant funding from global tech giants to be successful. https://rb.gy/2kyuyu

12

1 Comment

He gave a masterclass about responsible and trustworthy AI and LLMs.

22

2 Comments