Michelle Wang

Super interesting talk by Thomas Mullaney on the engineering challenge of inputting Chinese characters on a keyboard and how it was influenced by the intricate relationship between culture, technology, and geopolitics. I've been fascinated by this issue for a long time ever since I learned how to type Chinese with pinyin. One interesting difference between typing English and Chinese is that when you type English, every letter on screen is essentially WYSIWYG. But in Chinese, if you use pinyin for example the computer has to guess what characters you mean. It essentially becomes a search/predictive text problem, which is super interesting as we enter the age of LLMs which are just next token predictors. Can’t wait to read “The Chinese Computer”! #computerhistorymuseum

63

4 Comments

Day 22 of Learning in Public🚀 I've recently explored a groundbreaking paper on semi-supervised semantic segmentation, titled "CorrMatch," and I'm thrilled to share some transformative insights:   1. Efficient Label Propagation: Unlike traditional methods, CorrMatch introduces a novel framework that leverages correlation maps to utilize unlabeled data efficiently. These maps help in clustering pixels with similar characteristics, improving segmentation accuracy without requiring extensive labelled datasets.   2. Dual Propagation Strategies: The framework employs two key strategies:    - Pixel Propagation: This technique models the pairwise similarities between pixels, allowing high-confidence labels to propagate and enrich the segmentation predictions with global similarity information.    - Region Propagation: This method enhances pseudo labels by integrating accurate, class-agnostic masks derived from correlation maps, capturing object shapes effectively and expanding high-confidence regions.   3. Outstanding Performance: CorrMatch achieved a remarkable 76%+ mIoU score on the Pascal VOC 2012 dataset with only 92 labelled images. This performance underscores the potential of semi-supervised learning to produce high-quality results with minimal labelled data.   4. Streamlined Framework: One of the standout features of CorrMatch is its simplicity. Unlike other methods that require multiple networks, training stages, or strong data augmentation, CorrMatch offers a straightforward, single-stage process. This reduces computational complexity and makes the approach more accessible for real-world applications.   5. Comparison with State-of-the-Art: The paper highlights that CorrMatch surpasses existing methods like PS-MT, ST++, and UniMatch across various metrics. It achieves this by eliminating the need for multiple strong data augmentation streams and by simplifying the training pipeline.   6. Practical Implications: The insights from this paper are highly relevant for industries where labelled data is scarce or expensive to obtain. By improving the utilization of unlabeled data, CorrMatch can accelerate advancements in fields like autonomous driving, medical imaging, and environmental monitoring.   7. Future Directions: The framework opens new avenues for exploring the use of correlation maps in other machine-learning tasks, potentially extending its application beyond semantic segmentation to areas like object detection and image classification. 🔗 Explore the paper and code here: https://lnkd.in/gTanZS3g These insights have significantly broadened my perspective on the capabilities of semi-supervised learning. The simplicity and efficiency of CorrMatch make it a compelling tool for tackling complex segmentation tasks with limited data.   #MachineLearning #SemanticSegmentation #ArtificialIntelligence #DataScience #DeepLearning #Research #Innovation

4

That’s right.

2

🚀 Exciting News from Meta! 🚀 On April 18, Meta unveiled its latest breakthrough in AI technology – Llama 3, their largest generative AI model yet! 🎉 Outperforming even GPT-04 on certain standard AI benchmark tests, Llama 3 is poised to revolutionize the world of artificial intelligence. Named after its creators, Llama 3 is an LLM (Large Language Model) designed by Meta, boasting an impressive range of capabilities. From crafting creative writing pieces to coding assistance, brainstorming ideas to summarizing documents, Llama 3 shines in various applications. Perhaps most impressively, it can seamlessly respond in natural language to an array of queries, embodying the next frontier of AI interaction. But what sets Llama 3 apart isn't just its performance – it's also its accessibility. Meta has made Llama 3 open source, continuing their commitment to democratizing AI technology. This move ensures that developers and enthusiasts worldwide can leverage its power for their projects and innovations. Running Llama 3 locally has never been easier thanks to Ollama!(https://ollama.com/) 🦙 Ollama is an open-source tool that simplifies the process of operating LLMs on your own machine. With just a few simple steps, you can harness the full potential of Llama 3 right from your computer. Curious to learn more about Llama 3? Check out the official Hugging Face blog post (https://lnkd.in/gTysTg3E) for additional insights and details! Llama 3 comes in four variants: 8 billion parameters pretrained. 8 billion parameters instruction fine-tuned. 70 billion parameters pretrained. 70 billion parameters instruction fine-tuned. Let's explore the frontier of AI with Llama 3 and Ollama. And let's embark on this journey of endless possibilities together! 💡 #Meta #AI #Llama3 #Ollama #ArtificialIntelligence #OpenSource #Innovation #Technology #FutureOfWork

2

Here's a fun one for all of you: Recursively generating a prompt that generates flirtatious and NSFW messages. As part of Vellum.ai's offsite, I wanted to experiment with a different way of evaluating prompts: pitting two prompts against each other. Check out the full blog post for some cool learnings! https://lnkd.in/ezirpyWY

38

10 Comments

🚀 Excited to Share My Latest Project: Boston Housing Price Prediction! 🏠📊 I'm thrilled to announce that I’ve just wrapped up a comprehensive machine-learning project aimed at predicting housing prices in Boston. This project was a fantastic journey, where I got to apply a variety of regression models and enhance my data science skills. Check out the full project on GitHub: https://lnkd.in/da3xJjNV Key Highlights: 🔍 Exploratory Data Analysis (EDA): • Delved into the dataset to understand the key features impacting housing prices. • Visualized data distributions and correlations to uncover insightful trends. 🤖 Modeling: • Implemented a range of regression models including Linear Regression, Ridge, Lasso, SVM, KNN, Decision Tree, Random Forest, ExtraTrees, AdaBoost, Gradient Boosting, XGBoost, and MLP. • Evaluated models based on R2 Score and Mean Absolute Error (MAE). 🏆 Results: • The Stacking Regressor, combining Gradient Boosting, XGBoost, ExtraTrees, and RandomForest with a final Linear Regression estimator, delivered the best performance. • Achieved an R2 score of 0.973 and an MAE of 1.38. Skills & Tools: • Programming Languages: Python • Libraries: pandas, numpy, matplotlib, seaborn, scikit-learn, xgboost • Techniques: Data Preprocessing, Feature Scaling, Model Evaluation, GridSearchCV, Ensemble Learning This project not only sharpened my skills but also deepened my understanding of housing market dynamics. I'm always open to receiving constructive feedback and gaining new insights. #DataScience #MachineLearning #HousingMarket #Boston #Prediction #Python #Regression #EDA #DataAnalysis #DataAnalytics #BigData #ArtificialIntelligence #AI

13

10 Comments

I'm excited to share my latest video presentation on Cooperative Inverse Reinforcement Learning (CIRL) and its potential for aligning AI systems with human preferences. 🤖🤝👥 It's a bit scary for me to share this with my LinkedIn network, as this presentation was created as an assignment for my "ENGL 200 - Adv Composition & Technical Writing" course at Harrisburg University, and I don't post this type of content online.... In this video, I break down the key ideas behind CIRL and perform a SWOT analysis to examine it in the context of AI alignment. Watch the video to learn more about the pros and cons of CIRL and join the discussion on this exciting area of AI research! 🚀💡😎 #AIAlignment #CooperativeInverseReinforcementLearning #CIRL #AISafety

10

3 Comments

Finished the course on Prompt Engineering with Llama 2&3. Learned a lot about effective prompting techniques, which are valuable for both fine-tuning and improving model inference. https://lnkd.in/gEdqXMwH

16

Diving into AI: My Exciting Journey with Headstarter AI Fellowship! 🚀 Just wrapped up my first week of virtual meetings with the Headstarter AI team, and I'm buzzing with excitement! Though the fellowship officially kicks off on July 22nd, I've been soaking up knowledge daily from co-founders Yasin Ehsan 🚀 and Faizan Ahmed. Here's what we've covered so far: • Qualities of great software engineers • Networking tips for tech professionals • Standing out in the competitive SWE field I'm gearing up to build 5 AI projects over the next 7 weeks, culminating in a final project that I can't wait to share. Looking forward to collaborating with my team and attending some in-person meetups soon! What's your most valuable takeaway from a recent learning experience? Share in the comments! #HeadstarterAI #SoftwareEngineering #AIProjects #TechLearning 👨‍💻🌟

8

1 Comment

🌟 Excited to share a new project: Network of Trust 🌟 Earlier this year, I attended ETHDenver, where my friends Bhargavkumar K., Shawn Anderson, MSc., Dharmik Dholariya, and I participated in the hackathon. We built an impact attestation app for the Ecological Benefits Framework (EBF). Among over 200 projects, we placed in the top 15 overall, top 3 in the Impact & Public Goods track, and won several additional bounties, securing a total of $10K in prizes. Our success can be attributed not only to our technical prowess but also to our motivation to build for impact and contribute to the transformation of our economy. We aim to value the ecological foundations that our existence and economy depend on. The EBF, incubated by The Lexicon, is an amazing development in the ability to speak a common language of ecological and social impact that is emerging in a bottoms-up fashion to go above and beyond the existing frameworks of impact accounting. Now, we are building on our momentum and pursuing research and development through multiple pathways, and we would appreciate your support! We are honored to have been selected to participate in three Gitcoin public goods funding rounds that are running until May 7, 2024, 23:59 UTC: 1. Hackathon Alumni round: We are pursuing a collaboration with EBF and exploring adjacent applications for impact. Donate here: https://lnkd.in/gJ7wky2E 2. Token Engineering Commons round: We are building a peer-to-peer network to allow token engineers to attest to each other's skills. Donate here: https://lnkd.in/g8JP7m_6 3. Open Civics round: We are combining attestations with community knowledge networks to build tools for democratic civic collaboration at a local level. Donate here: https://lnkd.in/gBihQcau Donations to all of these rounds are matched via quadratic funding, meaning even a small donation counts as it is matched with an algorithm that favors plurality. If you have any questions about our project or how to contribute, please reach out! - Check out our project on Twitter: https://lnkd.in/g5Rk4BSr - Play with our app from the hackathon: https://lnkd.in/gawJHjKM - Explore our code: https://lnkd.in/gci3_Mgq #ETHDenver #Blockchain #Attestation #Ethereum #Impact #TokenEngineering #EBF

13

1 Comment

"Excited to share my latest work with Mercity Dive into the world of medical AI assistants with our new blog post: Building Medical AI Assistants with Visual LLMS. 🥼🩻🩺 Check it out here: https://lnkd.in/g_92K6cc 👨‍⚕️ #AI #HealthTech #MercityAI"

26

2 Comments

🏆 Exciting news! Our team, http418, emerged as the winners of the MongoDB Gen AI Hackathon with our project "tiktokify"! 🎥 Teammates Omkar Rajwade, Jeremy Herzog and I had an amazing time developing this tool, which automatically creates captivating 30-second highlight clips from videos using state-of-the-art embeddings and prompt engineering. With invaluable support from MongoDB, we turned our innovative vision into reality within just 8 hours. If you're curious to learn more about our project, check out my detailed blog post about our approach and experience at the hackathon: https://lnkd.in/d67DuiU5 A special thanks to Andriy Mulyar from Nomic AI for guiding us with their powerful embedding model and to Karissa Fuller for organizing such an exceptional event! 🙌 Explore our project code on GitHub: https://lnkd.in/dqGn3hbF If you're looking to hire or want to know more about our journey, feel free to connect with me on LinkedIn or have a look at my resume: https://lnkd.in/dvXrM6xW #Hackathon #AI #VideoSummarization #Innovation #Tech

374

34 Comments

Last weekend I wrote a new blog post reflecting on the prompt engineering techniques I used in a project called GARDEN (GPT Assisted Robust Development of Nanoservices) that I built about a year ago. Some key takeaways: - Chain of thought reasoning and breaking down tasks into smaller steps helps a lot. - Give the AI concrete examples to work from and detailed context. - Quality check the AI's outputs by having it evaluate its own work. - Context is king, but don't overload the prompt. "A key thing that went well is that instructions ended up at the end of the prompt, with all context above. Claude's docs recommend putting longform data at the top. Having instructions last, rather than before prior messages, really helps." Let me know what you think! https://lnkd.in/gCzHVMtu

7

Working on a technology you have been looking to work on for a while is always energizing. For me that tech was gRPC, and we recently leveraged it to develop a "Distributed File System" in the last project of our CS 6200 class. 🧑‍💻 The project was inspired by Andrew File System (One enterprise AFS deployment at Morgan Stanley exceeds 25,000 clients). The main task was to ensure that the files of all client nodes were consistent with the server and that the server was consistent with all of the clients within the distributed system for all files present in the system using remote procedure calls. 📂 We had to handle synchronization among two sets of threads, one for RPC calls to the server and another for Notification handling by the mounted client directories. ✅ In short, the project was an amazing introduction to coding distributed systems with numerous client nodes and a server node using multi-threaded and object-oriented programming in C++. 🙌 #grpc #distributedcomputing #multithreading #operatingsystems #computersystems #webservers #omscs

31

1 Comment

🚀 Exciting Summer Ahead! 🚀 I'm thrilled to share that this summer, I’ll be diving into a fascinating research project focused on evaluating state-of-the-art embedding models with instructions. The goal is to see if we can generate better embeddings that improve semantic representation. Here’s a glimpse of what I’ll be working on: 🔍 Using cutting-edge embedding models to process and understand complex datasets. 📚 Experimenting with domain-specific instructions to refine these embeddings. 📊 Leveraging datasets like DBLP and Amazon reviews, and using evaluation tools like the Hugging Face MTEB scoreboard. Why this matters: Embedding models are crucial for a variety of applications, from search engines to recommendation systems. By improving these models, we can make significant strides in how machines understand and process human language, ultimately leading to more intuitive and efficient technology. https://lnkd.in/ggF-3-fy #Research #MachineLearning #NLP #Embeddings #AI #Tech

2

📈 Excited to share my recent exploration into algo-trading and its fascinating world of data analysis! 🚀 Intrigued by the potential of algorithmic trading, I embarked on a journey to delve deeper into its mechanics. Over the past few weeks, I've immersed myself in the realm of stock market data analysis, particularly focusing on the concept of algo-trading. For those unfamiliar, algo-trading involves building buy or sell positions based on predefined conditions, often referred to as indicators in trading terminology. These conditions are crafted meticulously, utilizing a range of parameters to guide decision-making processes. My foray into this domain led me to develop custom indicators, tailored to generate trading signals based on specific criteria. To put these indicators to the test, I conducted a comprehensive study aimed at evaluating their performance. In one such experiment, I explored the performance of a custom indicator-based strategy, where buy and sell signals are generated based on certain conditions. To illustrate, consider a hypothetical scenario where a buy signal is triggered when Indicator A (say 20 days moving average) surpasses Indicator B (say 50 days moving average), and a square off signal is initiated upon the reverse. But here's where it gets interesting: rather than fixating on static parameter values( 20 or 50), I opted for a dynamic approach, testing a range of values to observe their impact on returns. The result? A simulated heatmap showcasing the fluctuation of returns over time across various parameter combinations. Each frame of the video represents returns over 45 trading sessions versus parameter values, with every frame change indicating a time period of 3 trading sessions. Reflecting on this experience, I've gleaned invaluable insights into the nuances of algo-trading: 1️⃣ Flexibility is key: Fixed parameter values don't guarantee consistent returns. What worked in the past may not necessarily yield the same results in the future. 2️⃣ Predictive challenges: Anticipating optimal parameters ahead of time remains a formidable task, given the ever-evolving nature of market dynamics. Looking ahead, I'm excited to explore avenues where this knowledge can be applied, such as high-frequency trading and the analysis of option strategies. By leveraging the insights gained from these experiments, we may uncover new opportunities to optimize trading strategies and enhance performance.

22

6 Comments

Out of curiosity: the math lineage of Jim Simons Jim Simons was shaped by a very solid lineage of mathematicians. Here, I use the sign < to denote a (doctoral) student-advisor relationship, considering only the main advisors of each researcher: Jim Simons < Bertram Kostant < Irving Ezra Segal < C. Einar Hille < Marcel Riesz < Leopold Fejér < Hermann Amandus Schwarz < Ernst Eduard Kummer < Heinrich Ferdinand Scherk < Friedrich Wilhelm Bessel < Carl Friedrich Gauß < Johann Friedrich Pfaff < Abraham Gotthelf Kästner < Christian August Hausen < Johann Christoph Wichmannshausen < Otto Mencke < Jakob Thomasius < Friedrich Leibniz. P.S.: Friedrich Leibniz was the father of the famous Gottfried Wilhelm Leibniz. RIP Jim Simons Source: The Mathematics Genealogy Project

22

3 Comments

🚀 Upon many requests across multiple platforms and as promised, today we're thrilled to launch our exclusive 25+ hour course "Fundamentals to Linear Algebra" - available exclusively on LunarTech app! 🚀 📚 Full semester of linear algebra in a focused course. Master the fundamentals in weeks, not months, and start building your own AI algorithms. 💪 🎓 Forget the PhD! You don't need years in academia to understand and create your own Machine Learning or Generative AI algorithms. What you do need is a solid grasp of the math combined with relevant industry insights, clear guidance to save your time, and a supportive community.🧠 🎯 This is a carefully crafted, detailed course combining academic rigor with industry application insights. It's designed to make your mathematics knowledge applicable both for acing undergraduate Linear Algebra exams AND advancing your career in Tech. ✨ Why Our Course is Unique: - Organized Learning: Starts with a clear roadmap of Linear Algebra for 2024, learn in structured and guided way using our LunarTech app! - Academic Rigor & Industry Insights while saving time: We combine the best of both worlds, ensuring your skills are relevant for both acing your exam and to level up your career. - Examples by hand: Every example is written from ground up by hand, with no assumptions made, ensuring you understand each concept thoroughly. 📘 Who This Course Is For: - Undergrad Students to ace Linear Algebra exams with our structured lessons from basics to advanced topics (this covers full academic semester of undergrad Linear Algebra). - Working Professionals to level up and dive into hashtag#machinelearning, hashtag#deeplearning, and hashtag#AI? This course will set you on the right path with applicable skills and insights. 📚 Course Prerequisites: - Essential Foundations:  A solid grasp of high school-level pre-algebra is necessary—think operations with variables, solving quadratic equations - Refreshers Included:  While we provide refreshments and detailed examples to jog your memory, having these basics will help you make the most out of your learning experience. - Calculus Knowledge:  Not required 🎉 Exclusive Access on LunarTech: For a limited time, this course is available as part of our Unlimited Plan, which also gives you access to many courses and our Ultimate Data Science Bootcamp, covering topics like machine learning, A-B testing, deep learning, and NLP, interview preparations etc. Note: This is a temporary offer as we plan to make this course a standalone due to its extensive content & high demand. 🌐 How to Sign Up: Use this link to subscribe to our plan. You'll also find details about other courses available. If you're ready to dive right in, you can skip the free trial and contact us directly for immediate access to the full course. Here is the courses Welcome Message! Sign Up and Start Your Free Trial: https://lnkd.in/e4Qq5xNi https://lnkd.in/ez4J8Kas

8