Steve Wright’s Post

Attention all Builders in Education and Research: Exciting news! Anthropic has launched the Claude 3.5 Sonnet model on Amazon Bedrock. This model offers even more intelligence than the Claude 3 Opus at one-fifth of the cost. Here's how the Claude 3.5 Sonnet enhances the experience for education customers: - Teaching assistance: Summarize long-form content, generate questions, create interactive quizzes, and evaluate student content. - Student support: Enhanced ability to understand context and nuance from complex tasks. Provides a more natural experience for chat-based learning. - Visual math processing: Effortlessly interpret and gain insights from intricate charts, graphs, and diagrams from math courses or research papers. Includes transcribing texts from imperfect images. - Expedited research: Summarize and generate high-quality statistical visualizations, actionable predictions, and valuable insights from unstructured research data. Stay ahead in the education sector with Anthropic's innovative Claude 3.5, Sonnet model, powered by #AWS! https://lnkd.in/dyFAyhWW #Education #Technology #Innovation

For readers aiming at a thorough understanding of Machine Learning: Allow me to share a noteworthy resource that has compelled me to share my thoughts on LinkedIn for the first time. ✨ 📘 "Machine Learning - A First Course for Engineers and Scientists" by Andreas Lindholm, Niklas Wahlström, Fredrik Lindsten, and Thomas Schön, published by Cambridge University Press in 2022. This book blends underlying mathematics with practical applications in machine learning and statistical modeling, covering important practical aspects such as cross-validation, model selection, and the bias-variance trade-off.  What sets this book apart for me is its capacity to spark a desire for further exploration while reading, adding a layer of comprehension that I was seeking as the authors skillfully demystify machine learning intricacies. It establishes itself as an invaluable tool for self-learners and those aspiring for deeper insight. For more details, visit 🔗 https://smlbook.org and explore the wealth of knowledge this resource has to offer. #machinelearning

A thought-provoking book Recently finished a book which highlights the harmful effects of mathematical algorithms used in AI – The author calls them Weapons of Math Destruction (WMD). Some points that are still hanging around in my mind from this book: - The three elements of WMD: Opacity, Scale and Damage. All of them will be present, to one degree or another for any example we consider. - They are opaque, unquestioned, and unaccountable, and they operate at scale to sort, target, or “optimize” millions of people. - Our own values and desires influence our choices; from the data we choose to collect and questions we ask. Models are opinions embedded in mathematics - When they are deployed into society, they become weapons of mass destruction because now they can be scaled quickly and efficiently. Author explains one by one so many areas and effects of these algorithms be it to assess the performance of teachers, college rankings, online ads, recruitment hiring, scheduling workers for operational efficiencies and so on. I read about so much work going on to make these algorithms transparent so their decisions can be reasoned about. But I think still a lot of them are still unregulated and unaccountable. Author mentions that unless the objective of using these algorithms does not change from merely for profit, these will continue to badly affect people on the other side of these algorithms and without owners of these algorithms feeling any guilt. https://lnkd.in/gTq6Q7uZ

2023 has been a year marked by lots of interesting books, and reads. Within them, I think that a special mention is deserved by Prof. François Fleuret's Little Book of Deep Learning, which is free to download for everyone (see link below). In this book, also available in a gently formatted phone-friendly edition, you find a primer for understanding the building blocks of DL without giving up on the necessary formalism. In a world where building can be easily short-cut by mere copy&paste + trial&error, taking time to understand the theory and methods behind code lines can be extremely enriching, and empowering. This was worth sharing, which is the first way to say thank you. https://lnkd.in/dc3eH568

Part of my learning journey this year: SVD - SIngular Value Decomposition What is it? Incredibly, every matrix can be factored using this technique, into 3 matrices. (not often in math you can say the words every, all, no exceptions). The 3 matrices are called left/right singular vectors and a diagonal in the middle of singular values. These are unique if the singular values are distinct. The SVD of a matrix X = USV* (U-left vectors, S diagonal singular values, and V-right vectors). Why do this? It helps to have a context for the matrix X - what is it? It could be a 10 megapixel photo, or a matrix of ratings for Hulu/Netflix of Users by Items with ratings or consuming data, or it could be a matrix of word embeddings. All three are complex data representations of the world, snapshots in high resolution perhaps...millions of pixels from an iphone photo, millions of movie ratings, or documents converted to data for language models. What is amazing about this? The singular values in the diagonal allow capture of signal, using "signal to noise" loosely. So if you pick the first few, say 100 singular values out of thousands in that diagonal, they will be the largest, contain the most signal and define the best approximate version of that matrix. (no kidding, the best approximation of that size (rank)). So for image compression, X = USV* is approximately U'S'V*', and if the S' is the first 100, this is the best approximation of the matrix by that size of matrix (rank 100). So - a 90% compression (e.g. from 2000x1500 to 100x100) of an iphone image with 100 singular values may look almost as crisp as the original photo. But it will be the best 100X100 compressed photo. Even more amazing, the SVD for a sparse matrix works well - when X is a ratings or watching matrix, it is very sparse as each person only watches a few movies out of the library. Even with all those missing values, the SVD truncated at a selected rank approximates well and reduces the computational nightmare of the large matrix. So - what is SVD? A matrix factorization into 3 parts (USV*)...that provides an ordered diagonal of singular values, allowing us to compress while retaining the bulk of the info in the original. This is a best approximation of the rank selected and we can dial it in to be as accurate as we desire. For a photo, it can provide compression or even watermarking (we can add in a watermark at the compressed level by modifying the singular values/vectors). For a ratings matrix it can be used to predict ratings and consumption patterns for unwatched movies or word patterns and sets the stage for the concepts that drive language models that are all the rage today (conceptually a start here only). Oh well - a start of a new concept. Let me know if I did not lay this out accurately - there are also cool nuances like latent factors I did not explore. The Python code is simple - just a line or 2 and wallah! SVD done. :) #Datascience #AIML #mathisamazing

Inquiry-based learning is more than using a search engine, and it's important to know the difference. One builds knowledge, the other not so much. #google #generativeAI #questions #futureoflearning #inquirybasedlearning https://lnkd.in/giH-EHF4

Very happy to see this chapter, which I co-authored with Kristian Bondo Hansen, available online: "The Use and Promises of Machine Learning in Financial Markets: From Mundane Practices to Complex Automated Systems" in Christian Borch and Juan Pablo Pardo-Guerra's (eds.) new Oxford Handbook of the Sociology of Machine Learning. Will be of interest to anyone interested in how machine learning is transforming the cultures of financial modelling and bringing about a closer alignment with Big Tech. https://lnkd.in/ebANt-tC

Take a look at NDIT’s K12 AI guidance!

Tuesday Evening Rant: Memorization. A skill that seems to be undervalued with progress in information and technology. Here are some thoughts my restless mind brought up at 2am last night. 1. Million of muslims around the world memorize the entire Quran, and billions memorize at least some parts of it. This has served as an important means to preserving the sacred text since the time of Prophet Muhammad (Pbuh). As a consequence, even if all the physical copies of the Quran disappeared, the entire text can be reproduced. Similarly, if someone were to try to corrupt the text, it would be easy to catch and correct such attempts. 2. Can say the same about mathematics? Do we have enough Haafidhs of Math so that we would still be able to rediscover math if all the textual knowledge was lost? Perhaps the answer is closer to "Yes" than we think, considering numerous experts who lecture on their specialties without notes. 3. A modern trend is that many instructors rely heavily on presentation slides in classrooms, impacting students' engagement with content. Worse still is the students' reliance on photos over note-taking, hindering long-term memory retention. There are no thoughts given to the transcription, no attempt to check the teacher's work or fill in the gaps. Consequently our devices get smarter, but not us. 4. Times tables. The debate over times tables and rote learning persists amid technological advancements. The paradigm shift brought around by calculators and smartphones makes learning times-tables ineffective. However, most of us don't actually know how to properly use a calculator. While calculators and smartphones alter learning methods, the efficiency of mental arithmetic remains unmatched. Computing simple products by typing on a calculator takes longer than accessing the memory synapse where your brain has stored the answer. 5. Math memorization should start earlier than times-tables. It should start with single digit addition and subtraction. You should not be computing 8 + 5 in your head, you should be retrieving that the answer is 13 immediately. Once you get into the habit, it shouldn't take long to start enjoying "knowing" more math than your peers by memorizing facts. 6. While understanding definitions and theorems is crucial, memorization plays a vital role. Formula sheets for exams should serve as a knowledge check, not a primary resource. How do we get good at memorization? It's a skill that improves with practice, and math provides us a great platform to practice this skill. What are your thoughts on this matter? Have you ever had to memorize content in other subjects? Did you find any value in doing so? Finally, do you also have such thoughts crowding your brain when you are trying to fall asleep? Please share.