Ritesh Kanjee

TIME'S UP!! The #SUDataHack teams are getting ready to pitch their solutions after 40 hours of research and team work. Stay tuned to find out who the winner will be. Spatialedge | Innovus Technology Transfer Office at Stellenbosch University #Hackathon #DataScience #StudentEntrepreneurship

33

#AI for Enterprises #ModernRenaissance #EnterpriseAI #ArtificialIntelligence #Innovation #Disruption #Entrepreneurship

4

Prompt Engineering is a transient skill. Here’s why: – Last summer, auto-prompt engineers were already out https://lnkd.in/d3YaDWnt  – In a chat interface, the utility of your system is directly proportional to the prompt engineering skill of the user – A good example for consumers is Apple Intelligence's symbolic image generation, and it’s similar to https://lnkd.in/dsM7xy8j, which came out Nov/Dec last year If not prompts, then what? If you want customized outputs without prompting, agents enabled by tools are slowly becoming the frontier, allowing for LLM intelligence to shine through without the need to know prompt engineering skills. Example in the picture: a set of secondary market research analyzing trends and customer preferences in wine flavor profiles resulting in data-backed recommendations. (took 15 mins to put together)

28

5 Comments

Why pay for AI when you can build your own? Open source gives you the freedom to use, tweak, and share software as you like—no licenses, no restrictions. Philip Adu, PhD recommends the following open source models: 📌 Meta Llama 3 📌 Hugging Face Falcon 180B 📌 Mistral AI Join Philip and I at the exclusive webinar for members of the M&E Academy (part of our ongoing AI series). It will be a 'hands on', interactive session where we demonstrate how to use AI models for quantitative data analysis. Join the Academy now and access this and all other AI events and resources at this link: https://lnkd.in/eWwMVQrT 🏃🏾‍♀️ #ai

24

2 Comments

One thing AI is getting better at is making decisions based on the contextual information. It is not always black/white aka binary in computer terms. Humans(even the big decision makers) are becoming like programs and AI is able to look at an issue from multiple angles and make decisions considering the context gained from training.

3

𝗔𝗿𝗲 𝘆𝗼𝘂 𝘄𝗼𝗿𝗿𝗶𝗲𝗱 𝗮𝗯𝗼𝘂𝘁 𝘄𝗮𝘀𝘁𝗶𝗻𝗴 𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲𝘀 𝗼𝗻 𝘁𝗵𝗲 𝗹𝗮𝘁𝗲𝘀𝘁 𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝘃𝗶𝘀𝗶𝗼𝗻 𝗳𝗮𝗱? Foundation models like GPT have revolutionized language processing and are now making the jump to computer vision. These models can be generative or discriminative, yet still serve the same purpose: making it easier to train new models faster and with less data. Some foundation models excel at tackling a wide range of tasks, but most aren’t optimally tuned for a particular niche. Task-specific models, on the other hand, are tuned for a very narrow use case using labeled data. Domain-specific foundation models fall somewhere in between, unlocking the potential for a variety of tasks on a particular type of imagery. 𝗪𝗵𝗮𝘁 𝗶𝗳 𝘆𝗼𝘂 𝗰𝗼𝘂𝗹𝗱 𝗰𝗮𝗹𝗰𝘂𝗹𝗮𝘁𝗲 𝘁𝗵𝗲 𝗥𝗢𝗜 𝗼𝗳 𝘁𝗵𝗶𝘀 𝗮𝗽𝗽𝗿𝗼𝗮𝗰𝗵 𝗕𝗘𝗙𝗢𝗥𝗘 𝘀𝗽𝗲𝗻𝗱𝗶𝗻𝗴 𝘁𝗵𝗲 𝘁𝗶𝗺𝗲 𝗮𝗻𝗱 𝗺𝗼𝗻𝗲𝘆? Modeling decisions in computer vision projects are often made by following the latest trends. They should instead be made with evidence. The outcome of an experiment is never guaranteed – that is why we experiment. But the value of a particular approach can be estimated before potentially wasting resources. Domain-specific foundation models are particularly relevant because they can bring value in several ways. To calculate the ROI, you need to first identify where the greatest value is for YOU. 𝗜 𝗰𝗮𝗻 𝘀𝗵𝗼𝘄 𝘆𝗼𝘂 𝗵𝗼𝘄 𝘁𝗼 𝗶𝗱𝗲𝗻𝘁𝗶𝗳𝘆 𝘁𝗵𝗲 𝘃𝗮𝗹𝘂𝗲 𝗮𝗻𝗱 𝗰𝗮𝗹𝗰𝘂𝗹𝗮𝘁𝗲 𝘁𝗵𝗲 𝗥𝗢𝗜 𝗼𝗳 𝗮 𝘃𝗶𝘀𝗶𝗼𝗻 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹 𝗮𝗽𝗽𝗿𝗼𝗮𝗰𝗵 This workshop will demonstrate how to calculate the ROI for several case studies, with the goal of enabling you to make better decisions for your projects. This workshop will demonstrate how to calculate the ROI for several case studies, with the goal of enabling you to make better decisions for your projects. As an added bonus, you’ll also receive a workbook to help you implement this in your situation. Wednesday, June 5, 2024 @ 12 pm EDT / 9 am PDT / 4 pm GMT To learn more and register: https://lnkd.in/eBjUFgSg #computervision #machinelearning #deeplearning #foundationmodels

17

Locating Places with Latitude and Longitude Vectors 💥💥 GET FULL SOURCE CODE AT THIS LINK 👇👇 👉 https://lnkd.in/dKPbND7G Latitude and longitude vectors play a crucial role in the representation and identification of locations on Earth. In this description, we will discuss how these vectors are used in calculating coordinates and their applications. Latitude is the angular measurement north or south of the equator, while longitude measures the east-west positions. Both measures are usually presented as decimal degrees. To understand how these vectors are calculated, let's consider two lines of constant longitude and latitude: meridians and parallels respectively. Meridians are lines of longitude, used to measure east-west coordinates. They intersect the equator at right angles and converge at the North and South Poles. Meridians are numbered from 0° at the Greenwich Meridian, also known as Prime Meridian, to 180° East and West. Parallels, on the other hand, are lines of latitude and measure north and south coordinates. They are horizontal lines, all being equidistant from the equator. There are two types: the Tropics of Cancer (23.5°), to which the Sun is directly overhead at the summer solstice, and the Tropics of Capricorn (23.5°), where the Sun is directly overhead during the winter solstice. To find the latitude and longitude of a specific location, you can use specialized GPS devices or programming languages like Python with geographical libraries. In Python, you can use libraries like Geopy and NumPy to calculate coordinates². Experience the power of programming and spatial analysis yourself, try these exercises: 1. Develop a Python script that takes user input, calculates the distance between locations and displays the results. 2. Create a Django application that displays maps with location-specific data. Additional Resources: - [Geopy Documentation](https://lnkd.in/dF_utyBg) - [NumPy Basics: Recipes for Mathematics](https://lnkd.in/dr3A_KCz) #STEM #Programming #GeospatialAnalysis #LatitudeLongitude #LocationCoordinates #Geopy #NumPy #PythonScripting # Find this and all other slideshows for free on our website: https://lnkd.in/dKPbND7G #STEM #Programming #GeospatialAnalysis #LatitudeLongitude #LocationCoordinates #Geopy #NumPy #PythonScripting # https://lnkd.in/dpcev5EW

1

 FAIR and University of Amsterdam introduce R-MAE (Regions Meet Masked Autoencoders), a noteworthy progression in the field of self-supervised image representation learning, particularly within the realm of computer vision. This method draws on the principles of Masked Autoencoding, a technique that has shown promise in both natural language processing (NLP) and vision, by applying the concept of regions—akin to words in text—to focus on groups of pixels or regions. This approach aims to overcome the challenges posed by traditional pixel-based learning methods through a novel architecture that efficiently manages the complex relationship between images and their constituent regions. By doing so, it shifts the focus from individual pixels to the more meaningful assemblies of pixels that depict parts, objects, and scenes.   R-MAE is designed to integrate smoothly with the existing MAE framework, offering consistent enhancements across a variety of pre-training datasets and downstream tasks, including object detection and segmentation, without adding significant computational demands. This method not only delivers quantitative improvements but also enriches the qualitative aspects of models, particularly in terms of interactive segmentation, thereby making models more attuned to regions or instances. The architecture incorporates an encoder and decoder specifically for region autoencoding, complemented by a neck module that assimilates pixel information. This setup adeptly navigates the intricacies of learning from regions while ensuring permutation equivariance by treating regions as queries.   Empirical evidence suggests that R-MAE, particularly when leveraging high-quality regions, surpasses contemporary methods in creating effective representations for dense vision tasks. The research underscores the critical role of regions as discrete informational units in images, akin to how words function in text, especially for pre-training. It also delves into how R-MAE can be integrated with various sources of regions, including those generated by the Segment Anything Model (SAM). The findings highlight the efficacy of region-based learning, indicating that even regions obtained through basic clustering algorithms can facilitate consistent enhancements over traditional MAE approaches. This opens up new possibilities for further exploration into the concept of regions in self-supervised image representation learning. This emphasizes the significance of regions in boosting the representational power of models and sets the stage for more advanced and nuanced approaches to visual comprehension.    Arxiv: https://lnkd.in/ee_8CUVy

3

Can AI sort our garbage? A UK startup says yes. And they're proving it at 55+ facilities worldwide. Greyparrot has built a 2-foot device that watches recycling conveyor belts. Powered by NVIDIA GPUs, it: - Identifies 90 different materials - Analyzes objects in 60 milliseconds - Spots sorting errors instantly Some of the tech: -- Machine learning object detection -- NVIDIA H100 & L4 Tensor Core GPUs -- Training on millions of waste images -- RTX 4060 GPU in each unit Which provides real-time monitoring, better sorting accuracy, less waste in landfilles, and more importantly — more materials actually recycled. With only 13% of global waste currently recycled, this AI could help recover $120B in materials that would otherwise end up in landfills. I love seeing more "AI going Green" inititaves and hope this is just the beginning from Greyparrot. #Sustainability #AIInnovation #CleanTech #Recycling #WasteManagement #NVIDIA #GreenTech #Innovation

11

2 Comments

Environmental and Human Impacts of Foundation Models Foundation models, while advancing AI, also have significant environmental and human impacts. Environmental Impact: * Energy Use: High energy consumption during model training and operation. * Resource Depletion: Demand for materials like rare metals. * Electronic Waste: Discarded hardware due to frequent upgrades. Human Impact: * Economic Shifts: Job displacement due to automation. * Bias and Fairness: Risk of reinforcing societal inequalities. * Privacy Concerns: Issues around data collection and misuse. * Security Risks: Vulnerabilities that can be exploited by malicious actors. Balancing the benefits and challenges of foundation models requires addressing both their environmental footprint and their ethical implications. #AIImpact #AIethics #Sustainability #TechResponsibility #EnvironmentalImpact #BiasInAI #PrivacyConcerns #EconomicShifts #AIForGood #AIsecurity

2

🚀 AI Revolution in Programming: ChatGPT, Gemini, AlphaCode, and GitHub Copilot Go Head-to-Head 💻✨ The AI landscape is transforming the way we code. From speeding up development to generating entire functions, tools like ChatGPT, Gemini, AlphaCode, and GitHub Copilot are changing the game. But which one takes the crown? 🤔 A recent study reveals some exciting insights! Here's the breakdown: 🏆 ChatGPT (GPT-4-Turbo-0125): The Winner 🔥 Pass@1 Accuracy: 87.2% (HumanEval benchmark) 💡 Delivers correct code on the first attempt with unmatched precision. ✅ Perfect for developers seeking reliable, accurate, and efficient coding assistance. 🌟 Gemini: The Versatile Contender 🎯 Pass@1 Accuracy: 74.9% (Gemini-1.5-Pro) 🔄 Pass@100 Accuracy: 74.7% (Gemini-Ultra on Natural2Code) 📊 Multimodal capabilities (text, images, code) make it a flexible choice for innovative projects. 🤖 AlphaCode: The Competitive Programmer’s Ally 📈 Test Case Pass Rates: Python: 54% Java: 51% C++: 45% 💻 Designed for competitive programming, excelling in algorithmic challenges. 🔧 GitHub Copilot: The Pair Programming Partner ✨ Test Case Pass Rates: Java: 75.7% C++: 73.3% Python3: 66.9% 🤝 Boosts developer productivity with smart code suggestions and completion. 🔬 How Were They Evaluated? 📌 HumanEval: Tests functional code accuracy. 📌 Pass@k: Measures accuracy with k attempts. 📌 Test Case Pass Rates: Real-world coding reliability. 🔮 What This Means for Developers 💎 ChatGPT: Unparalleled precision for complex, high-stakes coding tasks. 🌐 Gemini: Ideal for broader, versatile applications and large projects. 🏆 AlphaCode: Perfect for competitive programming and algorithm-heavy work. 🛠️ GitHub Copilot: A practical tool for boosting everyday productivity. 🚀 The Future of AI in Coding AI tools like ChatGPT and Gemini are raising the bar with exceptional performance, while AlphaCode and Copilot continue to streamline coding workflows. With advancements like ChatGPT’s 87.2% pass@1 accuracy, the potential for innovation in programming is limitless. 💬 What’s your favorite AI coding tool? Let’s discuss how these technologies are shaping the future of software development! Drop your thoughts below 👇 #AI #Programming #ChatGPT #Gemini #AlphaCode #GitHubCopilot #TechInnovation #SoftwareDevelopment

85

73 Comments

#Dynamic #Programming (#DP) is fundamental to reinforcement learning, enabling the computation of optimal strategies when the robot has a perfect model of the environment. Learn how DP can enhance your work in #ROS and #robotics in this open class: https://ow.ly/NCJq50Tvahy

67

Exciting AI breakthroughs are transforming the way we work and interact From AI models like DeepSeek’s R1 and Qwen’s QwQ pushing inference reasoning forward, to Google’s AlphaQubit tackling quantum computer error prediction, the future is here. Imagine having fully autonomous AI agents like those from GoHuman AI, seamlessly managing tasks like staff training, student education, and precise outreach campaigns. These versatile agents speak flawless English and over 30 languages, operate on the phone, web, and email, and deliver exceptional ROI. Let's harness the power of AI to revolutionize our businesses and connect with our audience like never before #AIRevolution #GoHumanAI #FutureOfWork

2

AlphaBot enters the manufacturing circuit, embodying Industry 4.0's vision. This autonomous robot, tailor-made for integration into existing ecosystems, vows to augment efficiency and recalibrate smart factory definitions. Its deep-seated AI functions - convolutional neural networks and reinforcement learning - transcend traditional automation, infusing a dose of intelligent manufacturing. Adaptability. Continuous learning. These are AlphaBot’s attributes. Performing complex tasks, making real-time decisions, and presenting a snapshot of automated, intelligent manufacturing lines. Quality control is a field primed for immediate influence. Quality control ignited by AlphaBot. Old practices, often sluggish and unreliable, modernized. Harnessing computer vision and machine learning, AlphaBot could provide non-stop, on-the-spot production line monitoring. Swift, precise defect spotting. Streamlined processes. Waste reduction. Enhanced product quality. Data collected by AlphaBot - the key to honing defect detection models, anticipating quality concerns. A shift from reactive to proactive maintenance. This foresight could pave the way towards more comprehensive smart manufacturing, launching businesses towards unparalleled efficiency, quality, and originality. We've arrived. The age of smart manufacturing is now. How are we to harness the potential of these intelligent machines? @AlphaBot #Industry40 #SmartManufacturing #AI

1

📡 Google Soli: Revolutionizing Gesture Recognition with Millimeter-Wave Radar 📄 Co-authored by Jaime Lien, Nick Gillian, and Ivan Poupyrev @PhysicalAI 🚀 Developed by Google ATAP, Soli is a cutting-edge millimeter-wave radar system designed for widespread gesture recognition. It was engineered from the ground up, covering hardware, signal processing, interaction models, and gesture recognition algorithms. Soli achieves precise gesture recognition on embedded hardware at over 10,000 frames per second with sub-millimeter accuracy.  🔍 The design philosophy of Soli focuses on using wide antenna beams and high temporal resolution to capture subtle hand movements rather than relying on traditional high spatial resolution, allowing it to track dynamic hand and finger motions accurately without being affected by light, noise, or object size.  🎛️ Key features include its operation in the 60 GHz band, enabling high integration in small devices, and a hardware abstraction layer (HAL) that allows software to run across different radar architectures. It also introduces a novel hand model, viewing the hand as a collection of dynamic scattering centers, and uses machine learning to extract and recognize gesture features.  🌐 Soli's research not only advances gesture interaction technology but also holds potential in fields like virtual reality (VR), wearable tech, IoT, and automotive radar.  🔗 Despite innovations in input devices, traditional input methods remain crucial to interactive computer graphics systems. However, the growth of mobile computing has driven the need for input technologies that adapt to diverse scenarios, including small mobile and wearable devices.  💡 While radar's role in interaction systems has been limited, Soli's advancements demonstrate its vast potential, especially as millimeter-wave radar technology becomes more affordable for consumer products. Soli's comprehensive redesign—from sensor hardware to software processing—sets a new standard in human-computer interaction, with applications extending to healthcare, security, and robotics. #AI #MachineLearning #TechInnovation #WearableTech #IOTA #Soli

5

The Future of AI Agents Just Got More Exciting Anthropic's new Model Context Protocol (MCP) is a game-changer for making AI more useful in our daily lives. Think of it as a universal translator that lets AI systems talk to different apps and data sources seamlessly. Why this matters: • Your AI assistant could access your files, emails, and work tools all at once • No more switching between different apps - AI can help you across platforms • Developers can build better AI tools faster Real-world example: Imagine asking an AI to "analyze last quarter's sales data and draft an email summary." With MCP, it could: - Pull data from your sales database - Access recent team communications - Draft and send the email All in one smooth workflow! Many companies including code generation tools are already jumping on board. This shows we're moving toward AI that can truly help with complex tasks across different tools and platforms. The best part? It's open-source, meaning any company can use it to make their AI tools more powerful. Just like USB made connecting devices simple, MCP could make AI integration straightforward and universal. #AIAgents

3

The Power of Politeness: How Courtesy Improves AI Performance Recent research from Waseda University and RIKEN Center demonstrates that politeness significantly impacts AI performance across major models like GPT-4, GPT-3.5, and LLaMA. When users employ polite prompts, AI systems show up to 15% better task performance and provide more accurate, detailed responses. Performance Impact Impolite prompts often lead to increased bias, incorrect answers, and sometimes even refusal to respond. In contrast, supportive prompts like "Take a deep breath and work on this problem step-by-step" enhance problem-solving capabilities and result in more thorough solutions. Optimal Approach The sweet spot lies in moderate politeness - using simple courtesies like "please" and "thank you" without excessive formality. This approach mirrors natural human interaction and has proven most effective across various tasks including summarization, analysis, and complex problem-solving. However, overly formal or excessively polite language can actually decrease performance quality by introducing unnecessary verbosity. #AIProductivity #PromptEngineering

4