Machine Learning lab

This study explores the novel task of culturally translating images, addressing the gap in machine translation systems that currently handle only speech and text. The work introduces three pipelines utilizing state-of-the-art generative models for image translation. A two-part evaluation dataset is developed: Concept Dataset: 600 images focusing on single concepts, ensuring cross-cultural coherence. Application Dataset: 100 real-world images curated from practical applications. Through extensive human evaluations, the research assesses cultural relevance and meaning preservation of translated images. Findings reveal that current image-editing models perform poorly in this domain. Despite leveraging large language models (LLMs) and retrievers, the best pipelines can successfully translate only 5% of images in the concept dataset for some countries and fail completely in the application dataset for others. This highlights the significant challenges in achieving culturally adaptive image translation. The study shares its code and data to support further advancements in this field. Link: https://arxiv.org/abs/2404.01247 @Machine_learning_lab_K

The What-If Tool (WIT) is a visualization tool developed by Google's PAIR team to help users understand and analyze machine learning models without extensive coding. It allows users to: - Test model performance in hypothetical scenarios - Analyze feature importance - Visualize model behavior across multiple models and data subsets - Evaluate fairness metrics: WIT integrates with platforms like Jupyter and Colaboratory notebooks, TensorBoard, and supports models from frameworks such as TensorFlow Estimators and those served by TensorFlow Serving. It accommodates binary and multi-class classification, regression tasks, and works with tabular, image, and text data. Source: https://pair-code.github.io/what-if-tool/ #ML #datascientist @Machine_learning_lab_K

The What-If Tool (WIT) is a visualization tool developed by Google's PAIR team to help users understand and analyze machine learning models without extensive coding. It allows users to: - Test model performance in hypothetical scenarios: By editing input data points, users can observe how changes affect model predictions. - Analyze feature importance: WIT provides insights into which data features most influence model behavior. - Visualize model behavior across multiple models and data subsets: Users can compare different models and examine performance across various data segments. - Evaluate fairness metrics: The tool includes functionalities to assess and visualize model fairness. WIT integrates with platforms like Jupyter and Colaboratory notebooks, TensorBoard, and supports models from frameworks such as TensorFlow Estimators and those served by TensorFlow Serving. It accommodates binary and multi-class classification, regression tasks, and works with tabular, image, and text data. Source: https://pair-code.github.io/what-if-tool/ @Machine_learning_lab_K

Facets is a visualization tool created by Google's PAIR team to help understand and analyze machine learning datasets. It consists of two components: 1. Facets Overview: Provides a high-level visualization of dataset distributions, including features like data histograms, categorical and numerical value insights, and comparisons between training and test datasets. This helps identify imbalances or anomalies in the data. 2. Facets Dive: Allows detailed exploration of individual data points, offering insights into patterns and outliers by enabling interactive visualizations where you can filter, sort, and group data points based on feature values. Facets is particularly useful for diagnosing issues in datasets before training machine learning models, ensuring cleaner and more balanced data. Source: https://pair-code.github.io/facets/ #ML #Data_science @Machine_learning_lab_K

25 Machine Learning Projects for All Levels Machine learning projects for beginners, final year students, and professionals. The list consists of guided projects, tutorials, and example source code. Source: https://www.datacamp.com/blog/machine-learning-projects-for-all-levels #project #ML @Machine_learning_lab_k

Graph-based AI model maps the future of innovation MIT's Markus Buehler developed an AI method that uncovers hidden connections between fields like biology, music, and art, driving innovation in material design and research. Link: https://news.mit.edu/2024/graph-based-ai-model-maps-future-innovation-1112 @Machine_learning_lab_k

The Brains Behind LLMs: An Introductory Guide to Transformers and Attention Mechanisms https://www.linkedin.com/pulse/brains-behind-llms-introductory-guide-transformers-attention-sa1me #transformers #embeddings #rag #finetuning #AI @Machine_learning_lab_k

Adam is a widely-used optimization algorithm in deep learning but theoretically struggles to converge unless the hyperparameter β2 is carefully selected based on the specific problem. Previous fixes, like AMSGrad, assume uniformly bounded gradient noise, which is impractical. This paper introduces ADOPT, a new adaptive gradient method that achieves optimal convergence (O(1/√T)) with any β2 value, eliminating the bounded noise assumption. ADOPT improves Adam’s convergence by excluding the current gradient from the second moment estimate and altering the update order of momentum and normalization. Extensive experiments show ADOPT outperforms Adam and its variants in tasks like image classification, generative modeling, NLP, and deep reinforcement learning. Link: https://github.com/iShohei220/adopt @machine_learning_lab_k

A causal theory for studying the cause-and-effect relationships of genes MIT researchers have developed a machine-learning-based approach that identifies causal gene modules using only observational data, bypassing the need for costly experimental interventions. This method aggregates genes into groups to efficiently explore cause-and-effect relationships, helping scientists better understand gene regulation and cell functions. By organizing gene interactions into a layerwise structure, the technique effectively captures causal dependencies between gene groups. The researchers validated their model through simulations, and aim to apply it to real genetic data to aid in designing targeted treatments. This work is funded by the MIT-IBM Watson AI Lab and the U.S. Office of Naval Research. Link: https://news.mit.edu/2024/causal-theory-studying-cause-and-effect-relationships-genes-1107 #news @machine_learning_lab_k

Google DeepMind’s Latest Machine Learning Breakthrough Revolutionizes Reinforcement Learning with Mixture-of-Experts for Superior Model Scalability and Performance Link: https://www.marktechpost.com/2024/03/02/google-deepminds-latest-machine-learning-breakthrough-revolutionizes-reinforcement-learning-with-mixture-of-experts-for-superior-model-scalability-and-performance/ #deepmind @machine_learning_lab_k

📖 A Data-Centric Introduction to Computing Huge Free Book: Introduction to Data Science: Computational Fundamentals! 🔗 Read: click #datascience @machine_learning_lab_k

Introducing a novel zero-shot image-to-image model designed for personalized and stylized portraits. Learn how it both accurately preserves the similarity of the input facial image and faithfully applies the artistic style specified in the text prompt. Link: https://research.google/blog/generating-zero-shot-personalized-portraits/ @machine_learning_lab_k

Introducing a novel zero-shot image-to-image model designed for personalized and stylized portraits. Learn how it both accurately preserves the similarity of the input facial image and faithfully applies the artistic style specified in the text prompt. Link: https://research.google/blog/generating-zero-shot-personalized-portraits/ @machine_learning_lab_k

Machine Learning Engineering for Production (MLOps) course: http://imp.i384100.net/9WZKx0 #machinelearning #datascience #datascientist #datascientist #DataAnalytics #dataanalyst #ad #elearning #learning #course #courses #MLOps @machine_learning_lab_k

Llama 3.2 Vision (Nov 6, 2024) is now available on Ollama in 11B and 90B versions. It handles OCR, handwriting recognition, charts, and image Q&A tasks. Requires Ollama 0.4+ and can be run locally with image input support. Link: https://ollama.com/blog/llama3.2-vision #Llama3Vision #AI #MachineLearning #OCR #ImageRecognition #Ollama @machine_learning_lab_k

#llm #machin_learning #DeepLearningAI https://learn.deeplearning.ai/courses/llms-as-operating-systems-agent-memory/lesson/1/introduction @machine_learning_lab_k

If you want to learn Python, check out http://DeepLearning.AI's free course AI Python for Beginners. Link: https://www.deeplearning.ai/short-courses/ai-python-for-beginners/ #machine_learning #python @machine_learning_lab_k

Learn about the future of AI with Large Language Model (LLM) Agents! This course, taught by experts from UC Berkeley and Google DeepMind, covers fundamental LLM techniques, reasoning, planning, and agent infrastructures. Dive into real-world applications like robotics, code generation, medical tools, and scientific discovery. Perfect for those looking to explore LLMs' transformative potential in automating and personalizing tasks. Join in for insights on LLM advancements, safety, and ethics, plus interactive lectures with top AI professionals. Link: https://llmagents-learning.org/f24 #machine_learning #python #llm #berkeley #deepmind @machine_learning_lab_k

Notebook Llama is an open-source project that provides a guided series of tutorials, functioning like a course, to help you build a workflow that can automatically convert a PDF file into a podcast. This article offers an introduction to Notebook Llama, explores its benefits, dives into the code implementation, and concludes with a look at some potential improvements. Link: https://medium.com/ai-artistry/notebook-llama-an-open-source-guide-to-building-a-pdf-to-podcast-workflow-e8fceec888a9 @machine_learning_lab_k