Mike Kwong’s Post

Manifold Learning, Recently, I have been reading about manifold learning, and I find it interesting that nonlinear manifold methods preserve the local relations between points. This is somewhat similar to using CNNs. Finding an appropriate manifold for data can significantly reduce the computational cost! Thanks to César Sánchez for motivating me to learn about it.

🚀 Exciting Update! 🎓 Just completed an advanced learning algorithm course for machine learning on Coursera! 🌐🧠 Grateful for the incredible insights and knowledge gained. Ready to apply these advanced techniques to real-world projects. Let's continue the journey of learning and innovation together! 💻🔍 I am also thankful to Dr. Jigar Sarda for your guidance and support and to Department of Computer Science and Engineering CSPIT, CHARUSAT UNIVERSITY for providing all facilities. #MachineLearning #Coursera #ContinuousLearning #TechEducation 🚀

CSAFE Learning March Webinar: Algorithmic Assessment of Striation Similarity between Wire Cuts Tuesday, March 19 • 1-2 PM CDT Presenter: Yuhang (Tom) Lin CSAFE graduate student Yuhang (Tom) Lin will present “Algorithmic Assessment of Striation Similarity between Wire Cuts” on March 19 at 1 p.m. CDT. Lin will propose a new, reproducible, automatic algorithm to analyze the similarity between wires, showing how interpreting the results promises a consistent way of using wires as forensic evidence in the future. Enroll for free at https://lnkd.in/dR3TP7in Attendees will have the chance to ask questions during the live presentation. But for those unable to attend, a recording of the webinar will be available on CSAFE Learning for later viewing. Webinar Description: The comparative evaluation of aluminum wire cuts holds considerable significance within the field of forensic science. Nonetheless, there exists an absence of a systematic algorithmic framework for assessing their degree of similarity. In our recently-introduced algorithm, we address this void by undertaking an examination of surface cuts presented in the x3p format. The outcome of this algorithmic procedure consists of cross-correlation values, which serve as measurements for the similarity. In order to enhance the precision of this process, a sequence of multiple procedural steps has been integrated into the algorithmic pipeline. These steps encompass various methodologies, including imputation, Hough transformation, and maximum likelihood estimation.

This certification signifies a comprehensive understanding of supervised learning algorithms, model training, and evaluation. The well-structured curriculum, and in-depth exploration of various algorithms have significantly enhanced my understanding of this dynamic field(Machine Learning).

How Crucial is Attention in Learning Regardless of the specific learning difficulty, attention plays a crucial role in the learning process. Think of it as the engine that drives your brain’s learning machine. It fuels various stages

Happy to share that our paper "Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning" got accepted to ICML2024! (https://lnkd.in/dgh32Jbb) In Multi-Task Learning it is common to first compute the gradient of each task and then aggregate all the gradients via some heuristic for updating the shared parameters. However, by only relying on the gradient values popular methods miss an important aspect, the sensitivity in each of the gradients' dimensions. Some dimensions may be more lenient for changes while others may be more restrictive. Hence, we propose a Bayesian view for this process. We place a probability distribution over the task-specific parameters, which in turn induce a distribution over the gradients of the tasks. As a result, we keep track of both the mean values and the variance (which reflect the sensitivity) in the gradients, allowing us to combine them more effectively! More details in the paper :-) Work done in collaboration with Idit Diamant, PhD, Arnon Netzer, GAL CHECHIK, Ethan Fetaya

Excited to share that I've successfully completed [introduction to machine learning] from Great Learning! A big thank you to everyone who supported me along the way. Looking forward to applying these new skills in my future projects. #ContinuousLearning #ProfessionalDevelopment #NewSkills

You're right! Hour-long first-person walking tour videos hold a lot of promise for self-supervised learning, and here's why: Natural Variations: As you mentioned, objects in the video undergo various transformations that mimic real-world experiences. They get closer (enlarged), further away (cropped), and appear from different angles (rotated). This continuous flow provides a rich training ground for an AI to learn robust representations of objects. Implicit Supervision: Unlike needing labeled data, the video itself offers inherent supervision. By observing how objects move and interact with the scene over time (following the laws of physics!), the AI can learn to predict their future positions or appearances. This allows the AI to develop an understanding of object permanence and relationships within the environment. Comparison to ImageNet: ImageNet, a massive dataset of static images, has been a workhorse for pretraining computer vision models. However, walking tour videos offer some potential advantages: Realism: Walking tour videos capture the natural dynamics of the world, including changes in perspective, lighting, and motion blur. This can lead to models that are better at generalizing to unseen data. Temporal Information: The video format provides a temporal dimension that static images lack. This allows the AI to learn how objects change over time and reason about their motion and interactions. Limitations to Consider It's important to note that while walking tour videos are promising, they might not completely replace ImageNet: Limited Scope: Walking tours might not cover all object categories as comprehensively as ImageNet. Specific object types or rare scenarios might be underrepresented. Focus on Ego-Motion: The first-person perspective inherently focuses on what's directly in front of the camera. This can limit the model's ability to learn about objects in the periphery or with different viewpoints. Overall, self-supervised learning from walking tour videos is a fascinating approach with significant potential. It leverages the richness of real-world experiences to train models that are more versatile and robust.

I’ve heard a lot of people say, “school na scam.” It means school is fraud. But is it? Today, I learned about the five steps of machine learning. They include: 1. Define the problem. 2. Build the dataset. 3. Train the model. 4. Evaluate the model. 5. Use the model. Today’s lesson focused more on defining the problem. I learned that you must identify a specific task first, and then determine the most suitable machine learning task to solve the problem. Then, the lesson introduced two types of machine learning tasks; supervised and unsupervised learning. To define supervised and unsupervised learning, I’ll have to introduce some terms like labeled and unlabeled data, #clustering, etc. I want this post to be as short as possible, so I won’t use so much #technical jargons. Moving on, while I was learning about labeled data, I was introduced to categorical and continuous (regression) label. This took me back to one of Engineer Chris’ digital #electronics class, where we had a little debate on what we thought discrete and continuous #data meant. Good memories😌 Anyway, it was easier for me to understand labeled data because of the #knowledge I had gotten from #school. Then it hit me. School may not really prepare you for the real world, but it makes it easier to understand certain real life situations related to your discipline. In fact, one of these machine learning steps involves #statistics. Things like standard deviation, mean, variance, range, and some of all those other stuff I thought wasn’t important in school. Now I get to use them to solve real life problems. And I’m certain it won’t be difficult for me to learn because it’s not foreign to me. In summary, I think I appreciate school more now. I appreciate my degree a little better. Do I think there’s room for improvements with our educational system? Definitely. My point is that it’s not entirely useless. What do you think? #machinelearning #physics #electronics #school #learning

🎉 𝗡𝗲𝘄 𝗙𝗿𝗲𝗲 𝗖𝗼𝘂𝗿𝘀𝗲 𝗔𝗹𝗲𝗿𝘁: 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗦𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻𝘀 𝗯𝘆 𝗨𝗖 𝗗𝗮𝘃𝗶𝘀 🎉 Unlock the power of theoretical computer simulations with this comprehensive course from UC Davis! Perfect for anyone looking to dive into agent-based models (ABM) and their applications in social science. 🔍 𝗪𝗵𝗮𝘁 𝗬𝗼𝘂'𝗹𝗹 𝗟𝗲𝗮𝗿𝗻: - Define Agent-Based Models (ABM): Understand what ABMs are and their significance. - Explore Schelling's Segregation Model: Analyze one of the most famous ABM examples. - Mix Different Models: Create new and complex simulations by combining various models. - Artificial Societies: Delve into sophisticated artificial societies like Sugarscape. - Practical Applications: Use computer simulations to solve real-world problems. - Capabilities & Limitations: Identify the strengths and weaknesses of ABMs. - NetLogo Coding: Learn to code and grow your own artificial society using NetLogo. 🕒 𝗖𝗼𝘂𝗿𝘀𝗲 𝗗𝗲𝘁𝗮𝗶𝗹𝘀: - Duration : 12 hours - Level : Beginner - Certification : Yes Don't miss this chance to elevate your skills and add a valuable certificate to your profile. 𝗘𝗻𝗿𝗼𝗹𝗹 𝗻𝗼𝘄 𝗳𝗼𝗿 𝗳𝗿𝗲𝗲 𝗮𝗻𝗱 𝘀𝘁𝗮𝗿𝘁 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘁𝗼𝗱𝗮𝘆❗ 🔗 𝗘𝗻𝗿𝗼𝗹𝗹 𝗛𝗲𝗿𝗲 : https://lnkd.in/gZcnbVH2 #ComputerSimulations #UCdavis #AgentBasedModels #SocialScience #ArtificialSocieties #NetLogo #OnlineLearning #Upskill #Certification