Active Learning Integration

High-quality, consistent annotations are essential for building robust machine learning (ML) models. However, conventional methods for training ML classifiers often require domain experts to annotate data, which is then passed to data scientists for model training, review, and iteration. This process can be resource-intensive and time-consuming. In a recent blog, Netflix's machine learning engineers share how they’ve developed a system to address these challenges. The business needed to create granular video understanding for various downstream applications, which required building ML classifiers capable of identifying visuals, concepts, and events within video segments. Their solution involves a three-step process to build these classifiers systematically. First, users (i.e. video experts) search for an initial set of examples from a large, diverse corpus to kickstart the annotation process. This is done through text-to-video search, powered by video and text encoders from a Vision-Language Model that extracts embeddings. Next, an active learning loop is used to build a lightweight binary classifier based on these embeddings. This classifier scores all video clips in the corpus and presents select examples to the user for further annotation and refinement. Finally, users review the fully annotated clips. This step helps spot annotation mistakes and discover new concepts, prompting users to revisit earlier stages for refinement when needed. This self-service architecture empowers video experts to continuously improve without relying on data scientists or third-party annotators. It has also demonstrated improved average precision over competitive baselines. With its multiple benefits, this system serves as a valuable reference. #machinelearning #datascience #activelearning #video #embedding #annotation – – –  Check out the "Snacks Weekly on Data Science" podcast and subscribe, where I explain in more detail the concepts discussed in this and future posts:    -- Spotify: https://lnkd.in/gKgaMvbh   -- Apple Podcast: https://lnkd.in/gj6aPBBY    -- Youtube: https://lnkd.in/gcwPeBmR https://lnkd.in/gU_9hFfN

Active Learning Strategies Active learning transforms students from passive listeners into active participants who question, apply, and connect their learning to real-world contexts. By engaging in doing, discussing, and creating, students retain knowledge more deeply, develop critical thinking and confidence, and see the relevance of what they learn. Collaboration with peers further builds empathy, teamwork, and essential lifelong skills beyond the classroom. The following strategies offer practical ways to bring these principles to life and help students actively engage with their learning. 💎 Students can have 2 minutes to prepare and gather their thoughts individually, then discuss in pairs for 10 minutes, before sharing perspectives with the class and having a class discussion. 💎 Students can have various roles to bring pro/con, or stakeholder perspectives to spark critical engagement. 💎 Students can be the “summarizer,” the “challenger,” or the “connector” (linking ideas to previous content), when it comes to group discussion. 💎 Students get a chance of extending conversations outside class by uploading their short 2-3 minute video reflection in the discussion forum. The video can include 3-5 key points or quotations from the resources that you brought to class, together with student reacting to them. 💎 Students present realistic scenarios and to solve or analyze them. 💎 Students act out decision-making situations (e.g., business negotiation, patient care, policy debate). 💎 After a mini-lecture, students get a 5-minute challenge where they can apply the concept to an example. 💎 Students create something tangible (a business plan, a design prototype, a policy brief) that has the key takeaways of the concept you taught. 💎 Students take short, low-stakes quizzes in groups where they remember and apply knowledge. 💎 Students individually or in a group teach a concept to the class and bring resources to support understanding. 💎 Each group learns one part of the content, then teaches it to others as a Jigsaw activity. 💎 Students make short videos, explainers, or infographics for presenting their findings to their peers. 💎 Students review each other’s work and provide constructive feedback, reinforcing their own understanding. What are some of the strategies that worked for your students?😊 #ActiveLearning #TeachingStrategies #StudentEngagement #DeepLearning #CriticalThinking #CollaborativeLearning #HigherEducation #InnovativeTeaching #LearningDesign #Pedagogy #EducationTransformation #LifelongLearning

Bayesian Active Learning to Efficiently Explore Physics-based Models Physics-based simulations, especially for complex phenomena like rapid solidification, are critical tools for understanding process-structure-property relations—but often remain exceedingly computationally expensive. Even with transformative advances in ML/AI-driven acceleration, fully exploring the different and complex regimes predicted by these simulations is a practical challenge. In this study–––carried out by our students José Mancias, Brent Vela, Juan Flores-Coronel and my co-authors, Rouhollah Tavakoli, Douglas Allaire, Damien Tourret––we directly address this key problem by integrating quantitative phase-field models with Bayesian active learning (BAL). Our goal: efficiently and cost-effectively explore and map critical microstructural transitions—such as dendritic-to-planar morphologies—in Fe-Cr alloys (a surrogate for 316L stainless steel under additive manufacturing conditions). 🧠 Key contributions include: Demonstrating the effectiveness of Bayesian active learning as a computationally efficient strategy for exploring multidimensional model input spaces in extremely expensive physics-based models (here, each simulation took multiple GPU-weeks to complete). Revealing and mapping previously unexplored intermediate, unstable microstructures in rapid solidification regimes whereby the transition region between planar and columnar growth is highly diffuse. Validating computational results against classical theories, confirming that the Bayesian-guided approach significantly accelerates the discovery of the location of important microstructural transitions. 💡 Broader Impact: Our approach paves the way toward practically leveraging costly physics-based simulations, providing a powerful and cost-efficient solution to systematically probe complex models. The framework is highly generalizable and can be deployed in any complex computational workflow whereby 'brute-force' model exploration is impractical due to computational costs. 🔗 Read the full paper here: https://lnkd.in/ednNVv9h #MaterialsScience #AdditiveManufacturing #MachineLearning #PhaseFieldModeling #RapidSolidification #BayesianOptimization #PhysicsBasedSimulations #Research

Google researchers just showed an active learning approach that slashed training data needs by up to 10,000× without hurting performance. The process sounds almost counterintuitive: • let the model flag the trickiest, most ambiguous cases • have human experts label only those • retrain on this curated, high-fidelity set For something like classifying unsafe ad content, where policies change often and the “right” answer isn’t always obvious, that’s huge. It means you can realign the model quickly, with far less cost and noise. They even used Cohen’s Kappa instead of accuracy, since there isn’t always a perfect ground truth in policy-heavy domains (smart choice). It rewards consistency between humans and models, not just correctness in some idealized dataset. And here’s the kicker: in their experiments, they went from ~100,000 training examples to under 500, while improving model–human agreement by more than 50%. Many ML projects fall into this trap. Teams spend time and money labeling easy, repetitive examples just because they’re easy to get. If a few hundred carefully chosen, expert-labeled examples can match or outperform the accuracy and alignment of a model trained on 100,000 randomly collected samples... why do we still default to thinking that more data will automatically make an ML model better? Check the research: https://lnkd.in/dtBgqXcD #ActiveLearning #MachineLearning #DataLabeling #ModelTraining #MLWorkflow #GoogleResearch

📖 Reading doesn’t have to be a silent, individual task. What if your students could teach each other while reading? Discover the Jigsaw Reading Method and other powerful strategies to make reading interactive and meaningful. Unlock Reading Skills with the Jigsaw Method, Students become “experts” on text parts, then share to build the full story. 2. Step 1 Break the text into smaller, manageable sections. 3. Step 2 Form expert groups—each group dives deep into one section. 4. Step 3 Reassign into mixed groups where students teach one another. 5. Step 4 Reconstruct the narrative together—everyone sees the bigger picture. 6. Helpful Supports ✔ Annotated passages ✔ Graphic organizers ✔ Guided prompts 7. Other Effective Reading Strategies ✨ Reciprocal Teaching – students take roles (summarizer, questioner, clarifier, predictor). ✨ Think–Pair–Share – quick reflection, peer discussion, then class sharing. ✨ Close Reading – multiple focused readings for deeper meaning. ✨ SQ3R Method – Survey, Question, Read, Recite, Review for structured comprehension. ✨ Reader’s Theater – dramatizing texts to build fluency and engagement. 8. Enhance the Experience 🔹 Use digital breakout rooms 🔹 Add visuals & timelines 🔹 Encourage peer questioning 🔹 End with a reflection or short writing task. Students don’t just read. They analyze, collaborate, and own the learning process. How do you make reading more interactive and collaborative in your classroom? #TeachingStrategies #ActiveLearning #ReadingComprehension #EnglishTeaching #EdTech #CollaborativeLearning #TeachingTips #StudentEngagement #LearningStrategies #OnlineTeaching

Active learning in ML explained visually. There’s not much we can do to build a supervised system when the data we begin with is unlabeled. Unsupervised techniques (if they fit the task) can be a solution. But supervised systems are typically on par with unsupervised ones. Another way, if feasible, is to rely on self-supervised learning. But it has limited applicability, which largely depends on the task and feasibility of "self-labeling." While full data annotation will work, it is difficult, expensive, and time-consuming. Active learning is a relatively easy, inexpensive, and quick way to address this. The visual below depicts how it works. The idea is to build the model with active human feedback on examples it is struggling with. There are 4 steps: Step 1) Manually label a tiny percentage of the dataset. Step 2) Train a model on this labeled dataset. This won’t be a perfect model, but that’s okay. Step 3) Generate prediction on the remaining unlabeled dataset and model's confidence. - We cannot determine if these predictions are correct as we do not have any labels. - That’s why we must use a model that can, either implicitly or explicitly, provide a confidence level about its predictions. - Probabilistic models (ones that output a probability of each class) are typically a good fit here. - One way to determine confidence is by looking at the difference between the top 2 class probabilities. - If the difference is large, this can indicate that the model is quite confident in its prediction. The opposite also holds true. Step 4) Label the lowest confidence predictions and feed them to the model with the seed data obtained in Step 1. Repeat this a few times and stop when you are satisfied with the performance. The only thing that you have to be careful about is generating confidence measures. If you mess this up, it will affect every subsequent training step. There's one more variant of active learning called cooperative learning. We covered it here: https://lnkd.in/gY2f_iAg. 👉 Over to you: What are some other efficient ways of building supervised models with unlabelled datasets?

Training a model for chemical reaction pathways often requires prior knowledge and expert intuition. A new active learning approach could reduce that burden. Machine-learned interatomic potentials (MLIPs) are trained on DFT data to approximate energies and forces across molecular configurations, offering a computationally efficient alternative to quantum methods in drug discovery and materials science. However, applying MLIPs to reactive systems remains challenging, where non-equilibrium events—bond breaking, bond formation, and transition states—are difficult to sample with conventional methods. Most training approaches depend on prior knowledge of products, reaction coordinates, or transition states. Recently published in JCTC, Siddarth Achar, Ph.D. et al. introduced a more automated approach: Reactive Active Learning (RAL). 🔹Bootstrap Setup: Begin with DFT-MD simulations of known reactants to train an initial committee of ML models (capturing model uncertainty). 🔹Reaction Exploration: Use this committee to propose and explore reaction pathways. Where models disagree most, flag those configurations as “uncertain” for further evaluation. 🔹Learning New Chemistry: Run DFT on newly discovered intermediates, then retrain the MLIPs to incorporate these species and pathways. 🔹Convergence Check: If uncertainty remains high, repeat the cycle until predictions stabilize. RAL allows models to discover reactions as they train. While limitations remain, such as barrier accuracy, the framework points toward a scalable route for building reactive models in complex chemical systems. 📄 Reactive Active Learning: An Efficient Approach for Training Machine Learning Interatomic Potentials for Reacting Systems, Journal of Chemical Theory and Computation, September 3, 2025 🔗 https://lnkd.in/enYn7KT6