Aymeric Roucher

💡 GPU? Lazy eval? No problem! 💻 Write dataframe code that works beautifully across pandas, Polars, cuDF & more with Narwhals. 🔥 Check out our in-depth blog, by Marco Gorelli on making data science more interoperable: https://lnkd.in/dedkSPPy #Polars #DataFrame #Pandas #Narwhals

14

📢 Already registered for #SDSC24? Don't forget to sign up for workshops! Check out all the sessions below👇 - Taming Big Data with #SpatialIndexing (hosted by CARTO) - Predicting demand using Snowflake Cortex ML and CARTO (Hosted by Snowflake) - Using Databricks and CARTO to process, serve and explore global spatio-temporal datasets (Hosted by Databricks) - Where and When: A Space-Time Analysis of Vehicle Collisions in London (Hosted by CARTO) - Risk and resilience: How well do you know your flood risks, today and tomorrow? (Hosted by Fathom) - Cosmopolitan Elites? Understanding the cost-of-living crisis in London at a hyper-local scale (Hosted by CARTO) 👀Want to sign up? Secure your spot for #SDSC24 for access to these workshops: https://lnkd.in/ePt9mpGD

25

1 commentaire

🔬 Exciting Breakthrough in AI Interpretability by Anthropic ! 🔬 Eight months ago, our team embarked on a challenging journey to scale sparse autoencoders for recovering monosemantic features from transformers. Today, we're thrilled to share a significant milestone: we've successfully extracted high-quality features from Claude 3 Sonnet, Anthropic's medium-sized production model! Our research has unveiled a fascinating diversity of highly abstract features that both respond to and influence behaviors. Some key findings include: - Features representing famous people, countries, cities, and type signatures in code. - Multilingual and multimodal features, bridging text and images across languages. - Abstract and concrete instantiations of ideas, such as code vulnerabilities and discussions on security. 📊 Key Results: 1. Interpretable Features: Sparse autoencoders are producing clear, interpretable features even for large models. 2. Scaling Laws: We've utilized scaling laws to guide our training, enhancing efficiency and effectiveness. 3. Abstract Generalization: The features generalize across multilingual, multimodal, concrete, and abstract references. 4. Systematic Relationships: A systematic relationship between concept frequency and dictionary size needed for feature resolution. 5. Influence on Behavior: These features can be used to steer large models, influencing behavior significantly. Importantly, the findings also touch on crucial safety concerns, revealing features related to deception, sycophancy, bias, and dangerous content. This breakthrough marks a significant step forward in AI safety and interpretability, paving the way for safer, more reliable AI systems. Kudos to the Anthropic interpretability team for this remarkable achievement! 🎉 Link to Paper: https://lnkd.in/ehH-eCdr #AISafety #MachineLearning #AIResearch #SparseAutoencoders #Transformers #Anthropic #Claude3Sonnet #TechInnovation #ArtificialIntelligence #AIInterpretability

5

In this Technical Deep Dive, we share our approaches in using Large Language Models (#LLMs) to make numerical-type data(sets) like DHS, JMP, HDX more accessible with the example of WASH data. In this 15 min. video Olivier Mills and team cover: 1️⃣ Basic scripting to extract core data from datasets from API endpoints 2️⃣ How to use LLMs to generate LLM-friendly schema definitions 3️⃣ How to structure a AI-friendly database store that can scale to thousands of datasets and millions of data indicators 4️⃣ How to use vector querying to retrieve relevant indicators from the store based off of a natural language (english) query 5️⃣ How to build a prompt that generates valid SQL query using LLMs with the above retrieved data schema. The various components form a single AI system combining data ingestion and data exploration using a software layer, a data layer and multiple LLM layers. ⏭️ Next steps will be to expand the datasets, automate other data source ingestion and further test the system with more complex plain english queries. We plan to incorporate this as a new feature for WASH AI after testing. Feel free to reach out if you would like a copy of the Notebooks, sample prompts and data schema definitions. We are #opensource The initial work was done in collaboration with Data for Good YYC (see https://lnkd.in/g3AijS93)

3

1 commentaire

🚀 Want to employ the power of #PyTorch for your scientific models without a steep learning curve? 👀 Check out our latest blog, where Travis Oliphant discusses how PyTorch's integration with the Array API makes it easy to use with familiar libraries like NumPy and scikit-learn.🔥 https://lnkd.in/dDRkEY3v #Innovation

10

🤖 MetaGPT Redefines Multi-Agent Collaboration 🤖 📄 arXiv 𝗣𝗮𝗽𝗲𝗿: "MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework" by Sirui Hong et al. During my talks, people usually ask me where I expect to see a boom in the incoming months. If Sam doesn't find AGI in the short term 😂, my answer is clear: Agents, Agents everywhere! So let's take a look to this paper about Agent Collaboration. 📝 𝗔𝗯𝘀𝘁𝗿𝗮𝗰𝘁 𝗦𝘂𝗺𝗺𝗮𝗿𝘆 Take a look to MetaGPT, a game-changer in multi-agent collaboration. By leveraging Meta Programming techniques, MetaGPT optimizes collaboration efficiency through Agent Roles, Workflows, Structured Messages, and Efficient Communication Protocols. Its outstanding performance on benchmarks like HumanEval and MBPP sets a new standard, particularly excelling in software development tasks. 𝗧𝗟;𝗗𝗥 🧠 MetaGPT introduces cutting-edge Meta Programming techniques for enhanced multi-agent collaboration. 📊 Achieves unparalleled performance on industry benchmarks like HumanEval and MBPP. 👨‍💻 Excels in software development tasks with high task completion rates. 🤝 Incorporates structured messages and efficient communication protocols for streamlined collaboration. 👾 Ablation study underscores the significance of diverse agent roles in boosting performance. 𝗙𝗼𝗿 𝗺𝗼𝗿𝗲 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 👇 📜 ArXiv Paper: https://lnkd.in/dt7XuM3Z 📪 Andrew Ng post on Multi-Agent Framework: https://lnkd.in/d5pmYcpB 👩‍💻 MetaGPT Github repository: https://lnkd.in/dQa7T4wM 👏 Kudos to Val Andrei Fajardo for the amazing visualization of the pager, it's pure magic for the eyes!

54

4 commentaires

Episode 3 of our AI/ML adventures: describing labeled Earth Observation (EO) data in a STAC index ! The use of the https://stacspec.org/ is crucial when it comes to describing spatio-temporal datasets, including labeled EO data. https://lnkd.in/et9B5GZg

6

𝐃𝐨𝐞𝐬 𝐲𝐨𝐮𝐫 𝐋𝐋𝐌𝐬 𝐬𝐩𝐞𝐚𝐤 𝐭𝐡𝐞 𝐓𝐫𝐮𝐭𝐡: 𝐄𝐧𝐬𝐮𝐫𝐞 𝐎𝐩𝐭𝐢𝐦𝐚𝐥 𝐑𝐞𝐥𝐢𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐨𝐟 𝐋𝐋𝐌𝐬 𝐰𝐢𝐭𝐡 𝐭𝐡𝐞 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐋𝐚𝐲𝐞𝐫 Modern Data 101 is here with its latest newsletter. In this edition, the authors delve into the challenges faced by current language model deployments and propose a solution through the semantic layer. By integrating a Semantic layer between the data and LLMs, organisations can enhance data comprehension, improve query accuracy, and reduce costs associated with data processing. Here are key takeaways: 💠 𝐄𝐧𝐡𝐚𝐧𝐜𝐞𝐝 𝐃𝐚𝐭𝐚 𝐈𝐧𝐭𝐞𝐫𝐚𝐜𝐭𝐢𝐨𝐧 𝐰𝐢𝐭𝐡 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐋𝐚𝐲𝐞𝐫𝐬: Implementing a semantic layer enhanced LLMs’ ability to understand and query data accurately, overcoming common issues like data inconsistency and context misunderstanding. This approach improves decision-making processes and reduces operational costs associated with data management and query optimization. 💠 𝐄𝐦𝐩𝐨𝐰𝐞𝐫𝐦𝐞𝐧𝐭 𝐨𝐟 𝐍𝐨𝐧-𝐓𝐞𝐜𝐡𝐧𝐢𝐜𝐚𝐥 𝐔𝐬𝐞𝐫𝐬: By integrating semantic layers, organizations empower non-technical teams to independently access and interpret data, accelerating business insights and improving overall operational efficiency without relying heavily on data scientists or analysts. 💠 𝐒𝐞𝐜𝐮𝐫𝐢𝐭𝐲 𝐚𝐧𝐝 𝐂𝐨𝐬𝐭 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧: Utilizing semantic layers for SQL generation ensures robust implementation of access control policies, mitigating security risks associated with direct data access. Additionally, it minimizes compute costs by optimizing data retrieval processes and maintaining data integrity within existing databases. 𝐑𝐞𝐚𝐝 𝐭𝐡𝐞 𝐞𝐧𝐭𝐢𝐫𝐞 𝐛𝐥𝐨𝐠 𝐡𝐞𝐫𝐞: https://lnkd.in/dR8gJbtN 📢 Do share what you feel about this take! 🔔 Follow Modern Data 101 and stay updated with our weekly highlights from the modern data space.

6

Missed our Anomaly Detection Live Deep Dive? You can watch it on demand. We had an amazing turnout and got great feedback on the session! Maayan Salom covered the following: - Data tests: assertions vs. anomaly detection - How anomaly detection tests work - How to configure anomaly detection tests (step-by-step) - Best practices Check it out here: https://lnkd.in/dwcPP2p5 We also recommend joining our anomaly detection channel on Slack. The link is available in the first comment. #anomalydetection #dbt #dataengineering

17

1 commentaire

In our latest w/ CASM Technology, we look at possible platform manipulation on Wikipedia, testing out a method called semantic clustering on the ENG-language Wikipedia entry for the Russo-Ukrainian war, & 48 other pages about Ukraine that link directly to it. By clustering the editors of these pages on the basis of the meaning (the ‘semantic content’) of the edits made, we identified 176 total clusters. While useful in grouping different accounts making very similar edits, it didn't necessarily prove if the edits were legitimate. We did however identify clusters that contained a number of single-controller ‘sock-puppet’ accounts. This typically happened when an editor intended to make a consistent series of changes, and either used multiple accounts to do so to circumvent a ban or to camouflage their activity. One cluster, for example, is composed of 9 editors & 4 sock-puppets, with the 2 most prolific linked to a single user. All edits from the sock-puppets are to the ‘Russia–Ukraine relations’ page & consist of updating stats on the country comparison table, including army sizes for each. In the future, semantic clustering for detection could help identify accounts with highly similar edits for closer scrutiny. However, any automated analysis should be linked to Wikipedia's human-led investigations & sanctions regime, such as Wikipedia Administrators's CheckUsers). As manipulation tactics evolve, empowering Wikipedia's large volunteer community with effective tools will be of high importance in order to conserve the platform's integrity. Our full paper below 👇 https://lnkd.in/dh3gxejs

28

Large Generative Graph Models (LGGMs): A New Class of Graph Generative Model Trained on a Large Corpus of Graphs Quick read: https://lnkd.in/gt_GbngB Paper: https://lnkd.in/gP27y6Gq

20

Accelerating ViTs 1.46x in PyTorch using our new Block Sparse Kernels 🔥 Check it out: https://hubs.la/Q02x72-c0

214

💥 Exciting update on accelerating vision transformers! See how the PyTorch team, Meta's FAIR team, and Quansight sped up ViTs using Block Sparsity. They demonstrated up to 1.46x speedup on ViT-B-16 using block sparsity and pruning on module weights with <2% accuracy drop. Key enabler: efficient sparse kernels in @PyTorch. Shout out to Andrew M. James, Pearu Peterson, Nikita (Nik) Vedeneev

7

2 commentaires

🚀 Some great bits of advice from the leaders at the VivaTech conference, Paris Meta AI chief Yann LeCun, advised students looking to work in the AI ecosystem, to not work on LLMs- “If you are a student interested in building the next generation of AI systems, don’t work on LLMs. This is in the hands of large companies, there’s nothing you can bring to the table” Some more key points to take away from the conference:- 1️⃣ LeCun asked people to focus on developing next-generation AI systems that overcome the limitations of large language models. 2️⃣ Muffed Vh the creator of Devika told people should move away from Transformer models and start researching on new architectures. He believes that with this approach, it is even possible to build something nearly as impressive as GPT-4 3️⃣ They also spoke about the growth of AI assistants and how all our interactions with the digital world will be mediated by AI assistants in the coming future. A great article on what we should focus on building next, to create an impact. Link - https://lnkd.in/dfb7g4PF #generativeai #ai #llm #openai #machinelearning #meta #analyticsindia

10

I am proud to be working at Martian, where our AI research team is pushing the boundaries of scaling Mechanistic Interpretability, arguably one of the most important areas of AI research. In recent weeks, we have seen a surge of progress in AI interpretability, driven by both Anthropic and OpenAI. While their recent techniques (linked in our blog post) involve deep and painstaking human inspection of individual neurons, our approach aims to automate this understanding at scale. The Martian research team, led by Shriyash U. and Chaithanya B., outlines a groundbreaking insight in this blog post: to truly automate the understanding of neural networks, we must evolve not only our techniques but also our theoretical foundations. They propose category theory as the key to this evolution. As our founder Etan Ginsberg would say, for more on the technical details, dive into the blog post. For those who are “normies” like me, you might wonder why this is so important! The goal of Mechanistic Interpretability is to make AI models understandable, so we can "read" them like a book. Currently, AI models are largely black boxes—even their developers are often surprised by what they can do. Understanding how to prompt them, what they excel at, and their inherent biases is a process of trial and error. Imagine the Possibilities: -Having detailed insights into an AI model's knowledge and capabilities. -Understanding its biases, strengths, and weaknesses. Knowing how to effectively prompt a model to get the desired answers. This level of understanding would revolutionize our interaction with AI, making it more transparent and predictable. At Martian, this research powers our LLM Router, which predicts the best large language model (LLM) for a given prompt based on a customer's cost, latency, and output quality preferences, a highly practical real-world application of Mechanistic Interpretability.  Follow Martian to be tuned into more exciting developments from our team! P.S. - As I write this, our founders just came out of a meeting with the research team of a major global tech company. They were delving into our technology, and the customer said, “You guys are better than OpenAI.” While I’m not sure of the exact context, it’s still incredibly flattering! #AI #MachineLearning #ArtificialIntelligence #TechInnovation #MechanisticInterpretability #Research #MartianAI #AITransparency

7

Foundation models are a hot topic right now, especially in domains like satellite imagery, where there is a lot of unlabeled data. But is this always the best approach? Yiqun Xie et al. questioned how a foundation model approach compares with supervised deep learning and more traditional methods like random forests and XGBoost on multispectral satellite imagery. They evaluated these approaches for the pixel-level tasks multi-temporal crop classification, wildfire scar, and flood mapping. The top-performing methods were a U-net, DeepLabv3+, and XGBoost, for the three datasets respectively. Clearly, naively applying a foundation model will not necessarily outperform other approaches. There's also a hybrid approach that they did not try: fine-tuning a self-supervised model pretrained on imagery similar to these datasets, rather than on a broad set of satellite imagery like the foundation models they tested. Sometimes this domain-specific approach, which could simply involve a second pretraining step with an existing foundation model, can provide a more powerful set of features and simplify the final supervised training. https://lnkd.in/eQps7xuK Want to learn more? Sign up for my Computer Vision Insights newsletter: https://lnkd.in/g9bSuQDP And if you need help training or fine-tuning your own domain-specific foundation model, please reach out. #RemoteSensing #EarthObservation #MachineLearning #DeepLearning #ComputerVision

75

4 commentaires

Nicolai Nielsen just released a great video about using Frame Interpolation in custom datasets for all machine learning enthusiasts! 🤖 You can learn about: 💡Setting up key frames in Keylabs to annotate a lot of data at the same time. 💡Team annotation for taking care of larger labeling tasks. 💡Various use cases for you to get started with your own projects. Check it out on YouTube! https://lnkd.in/dJ_9kezb #machinelearning #computervision

13

Who said RL doesn't generalize? 👉 Check out PWM, which builds on TD-MPC2. PWM can be deployed on new tasks (< 10m fine tuning) using pre-trained TD-MPC2. It's multi-task, it supports high-dimensional control tasks... what else could you be asking for? The one thing that really astonishes me with RL research today is how close the field is getting to more "consensual" ML: foundational models, offline datasets, fine-tuning... We're very far from the E2E one-fits-all solution that many people think of when they talk about RL. (Oh, and did I mention that they use tensordict and torchrl? 🤠 ) Work by Ignat Georgiev🤖 Nicklas Hansen and team Site: https://lnkd.in/eZGG63ym #reinforcementlearning #AI #robotics

72

3 commentaires