AI Community

VideoPrism: A Game Changer for Video Understanding Excited to share our latest research on VideoPrism, a powerful new model for understanding videos! Imagine unlocking the secrets hidden within millions of videos: from everyday moments to historical events and scientific observations. VideoPrism is here to make that possible! This groundbreaking ViFM tackles a wide range of tasks, including: - Classification: Identifying objects and actions in videos. ✅ - Localization: Pinpointing where things happen in videos. - Retrieval: Finding videos based on text descriptions. - Captioning: Generating captions that accurately describe videos. ✍️ - Question Answering: Answering questions about the content of videos. What makes VideoPrism unique? - Massive and diverse pre-training data: We trained VideoPrism on a huge dataset of videos and text, including high-quality captions and even noisy auto-generated transcripts. - Two-stage training: This innovative approach helps the model learn from both the visual content and the accompanying text, leading to a deeper understanding of videos. - State-of-the-art performance: VideoPrism outperforms previous models on a wide range of benchmarks, showing its exceptional capabilities. Beyond the benchmarks: - Combining with LLMs: VideoPrism can be paired with language models to unlock even more powerful applications, like video-text retrieval, captioning, and question answering. - Scientific applications: VideoPrism is already showing promise in assisting scientists in various fields, like ethology and ecology. Read more: https://lnkd.in/e4Zi9R73 #VideoUnderstanding #AI #Science #MachineLearning #Research

🚀 Take Privacy-Powered Typing Experience to New Heights with Gboard 🛡️✨ In the fast-paced world of technology, Google Research is setting new standards for privacy and efficiency with its latest advancements in Gboard's language models (LMs). Powering features like next word prediction and Smart Compose, these on-device LMs are enhancing the user experience while prioritizing data privacy like never before. 📱💡 🔐 With the integration of federated learning (FL) since 2017 and the addition of formal differential privacy (DP) guarantees in 2022, Gboard is leading the way in private, collaborative model training directly on users' devices. This groundbreaking approach ensures user data remains on the device, offering both lower latency and superior privacy protection. 🌍 🛠️ Innovative Privacy Measures in Action: - Transparency and Control: Gboard users are informed about how their data is used and have easy access to configure these settings. - Data Minimization: By focusing solely on updates that enhance model performance, FL ensures that only necessary data is processed. - Anonymization through DP: Applying DP during server processing prevents the model from learning unique user information, offering quantifiable privacy measures. - Auditability: Google's commitment to openness is evident in its public disclosure of algorithmic approaches and privacy accounting. 🌟 A Leap Forward in Privacy Guarantees: Gboard's on-device LMs, now deployed in over 7 languages across 15+ countries, adhere to stringent (ɛ, δ)-DP guarantees, marking a significant milestone in user-level DP deployment. The journey towards even stronger DP guarantees continues, with models for Portuguese in Brazil and Spanish in Latin America achieving an ε ≤ 1, setting a new standard for privacy in machine learning. Click here to read more: https://lnkd.in/djTmTJeY #GoogleResearch #Gboard #PrivacyInnovation #FederatedLearning #DifferentialPrivacy #AI #MachineLearning

Android Early Access Preview for On-Device Generative AI Gemini Nano runs on-device and enables features where the data should not leave the device, such as suggesting replies to messages in an end-to-end encrypted messaging app. https://lnkd.in/gDMB-jAR

Introducing Gemini: Google's largest and most capable AI model 🚀🌐 Google Research is at the forefront of a groundbreaking shift in AI technology, promising to revolutionize the digital world even more profoundly than mobile or the web. As they navigate their eighth year as an AI-first company, the pace of innovation isn't just sustained - it's accelerating exponentially! 🌟 Millions are already harnessing the power of generative AI in daily life, from complex problem-solving to creative collaborations. Google's AI models and infrastructure are not just transforming user experiences; they're fueling the growth of startups and enterprises globally. 🌍💡 What's truly remarkable is that we're just beginning to unveil AI's full potential. Google Research's commitment to bold, responsible research is paving the way for AI technologies that promise enormous benefits while incorporating safeguards to address evolving risks. Their investment in cutting-edge tools, models, and infrastructure, adhering to their AI Principles, is guiding this transformative journey. 🛡️🔍 Introducing Gemini, Google's most advanced and versatile model yet, marks a significant milestone in this journey. Gemini 1.0, comprising Ultra, Pro, and Nano versions, showcases state-of-the-art performance across multiple benchmarks. This innovation is a testament to Google DeepMind's vision and represents one of the most significant science and engineering efforts in Google's history. 🤖🔝 Gemini is more than just a model; it's a beacon of hope and opportunity, set to unlock new possibilities for people across the globe. With capabilities spanning text, code, audio, image, and video understanding, Gemini stands as a testament to the future of AI — a future where technology is not just a tool but a partner in advancing human creativity and productivity. 🌟🌏 The road ahead is filled with untapped potential and uncharted territories. Google Research's commitment to exploring these realms responsibly and collaboratively sets a precedent for the future of AI, ensuring that the advancements we make today pave the way for a more informed, creative, and productive tomorrow. 🚀🤝 https://lnkd.in/dAyySX2p #Google #GoogleAi #Gemini #AIRevolution #GoogleResearch #Innovation #Technology #GeminiModel #ArtificialIntelligence #FutureIsNow #GoogleDeepMind

🔬🧠 Exciting times ahead in the realm of neuroscience! Google Research is diving deep into the mysteries of the brain. Their latest project? Embarking on a mission to map a mouse brain. Why? Well, the human brain, a marvel in itself, is a complicated maze, and understanding how glitches in its network can lead to illnesses such as dementia remains elusive. Enter connectomics - a field dedicated to mapping every connection in the brain, potentially unlocking answers to our most pressing neurological questions. 🔍🔓 The Connectomics team at Google Research, in collaboration with leading institutions like Harvard University, the Allen Institute, MIT, Cambridge University, Princeton University, and Johns Hopkins University, is venturing into a whopping $33 million project over the next five years. With support from the BRAIN Initiative at NIH, the goal is audacious: mapping 2-3% of the mouse brain, targeting the memory-centric hippocampal region. 🐭📌 Remember the Human Genome Project's early strides? The same spirit drives this venture. Last year, a milestone was achieved when one cubic millimeter of the human brain was mapped, resulting in the H01 dataset. 🧬📊 With the mouse brain being a mirror into human brain functionality, this endeavor aims to create one of biology's largest datasets - a staggering 25 petabytes of brain data. To give you an idea, that's equivalent to 100,000 Milky Way Galaxies! 🌌🌌 And as the data pours in, Google Research is prepared, enhancing its cutting-edge tools and technologies, ensuring that these massive datasets are not only manageable but meaningful. 💡🛠 Exploring the intricacies of the mouse brain might just shed light on the wonders of our own. Here's to the promising journey ahead, where the micro offers a lens into the macro! 🌌➡️🔬 https://lnkd.in/g3RhZdcT #Neuroscience #BrainResearch #Connectomics #GoogleResearch #InnovationInScience

Recent studies have shown that while Large Language Models (LLMs) offer data-efficient learning, their size can pose real-world deployment challenges. Deploying a 175 billion LLM, for instance, requires a massive 350GB of GPU memory and specialized infrastructure! 😱 But, what if we could outperform these behemoths with smaller model sizes and less training data? Introducing "Distilling Step-by-Step" 🌪️✨, an innovative mechanism highlighted in the paper presented at #ACL2023. This approach allows for training smaller, task-specific models with significantly less training data than standard approaches, yet these models can outperform even the few-shot prompted 540B PaLM model! 🌐 How It Works: - Distilling step-by-step leverages informative natural language rationales from LLMs to train smaller models more efficiently. - These rationales explain connections between input and output. - The technique uses a multi-task problem framing to train the model with a new rationale generation task. 📉 Key Results: - Achieves better performance using up to 80% less training data on benchmark datasets. - A 770M parameter T5 model outperformed the 540B PaLM model, a more than 700x model size reduction! 📊 In Practice: With distilling step-by-step, not only is there a significant reduction in model size, but the amount of data required for training also sees a massive drop. This means we're looking at a future where high-performance language models are more accessible and feasible to deploy in real-world applications, even for smaller research teams! Cheers to the innovators behind this paradigm shift! 🥂 Reducing both deployed model size and the amount of training data required? Now that's what we call #AIRevolution! 🎉🔥 https://lnkd.in/gw2xavkd #LanguageModels #AI #Innovation #Distillation

🔥 Exciting news in the #AugmentedReality world! The rise of real-time face feature generation and editing in mobile apps is undeniable, and there's now a groundbreaking solution to meet the demand: MediaPipe FaceStylizer! 📲💡 MediaPipe FaceStylizer is an efficient design for few-shot face stylization, crafted to address model complexity and data efficiency challenges, all in line with Google’s responsible #AI Principles. 🌐🤖 🔍 Spotlight Features: 1️⃣ An auxiliary head that converts features to RGB, enabling high-quality images generation from coarse to fine granularities. 2️⃣ Unique loss functions combined with common #GAN loss functions. 3️⃣ Open-source availability through MediaPipe, allowing users to fine-tune with personal style preferences! 🎨 💡 How Does It Work?: The team developed BlazeStyleGAN, based on #StyleGAN, designed for efficient on-device face generation. This design significantly reduces the model's complexity while maintaining impeccable generation quality. Moreover, they've introduced an efficient GAN inversion, acting as the encoder, to map input images into the generator's latent space. This innovative approach supports a more personalized image-to-image stylization. 🌀✨ 📊 Performance Metrics: BlazeStyleGAN not only slashes model complexity but also excels in on-device performance. Real-time performance is achieved on numerous high-end mobile devices, including the latest iPhone 13 Pro and Samsung Galaxy S20! 📈📱 🌐 Fairness: Ensuring fairness in AI is crucial. The FaceStylizer has been trained on a diverse human face dataset, promising a balanced performance across genders, skin-tones, and ages. 🌍💙 🖼️ Visual Treat: The stylization results showcase the prowess of MediaPipe FaceStylizer, demonstrating impeccable face stylization across a range of popular styles. A blend of art and tech! 🎨🔧 🚀 The Future: The MediaPipe FaceStylizer is set to launch publicly via MediaPipe Solutions. Users can harness the MediaPipe Model Maker to train a custom face stylization model with their style preferences and deploy across various platforms effortlessly. 🌌🛠️ Kudos to the teams behind this innovation for seamlessly blending art and tech, and paving the way for the next-gen #AR experiences! 🎉 https://lnkd.in/gJu6Pk2D #FaceStylization #MediaPipe #TechTrends

🌍🛣️ Routing on #GoogleMaps just reached a game-changing milestone! Determining the ideal route involves intricate trade-offs, and to optimize this, real-world travel patterns are invaluable insights. 🚗🛵 Enter the world of Inverse Reinforcement Learning (IRL). With IRL, Google Maps is able to capture travelers' preferences by observing real journeys. While traditional methods of IRL face scalability challenges, a collaborative effort between #GoogleResearch, #Maps, and #GoogleDeepMind has unlocked a solution! 🚀🔍 By innovatively utilizing graph compression, parallelization, and a novel algorithm called Receding Horizon Inverse Planning (RHIP), they've managed to break through the IRL scalability barrier. The result? A whopping 16-24% improvement in the route match rate globally! This implies that Google Maps' route suggestions now align more closely with users' real-world choices. 🌏📍 For the tech enthusiasts: RHIP borrows inspiration from humans' planning tendencies. Think about how we make detailed plans for the immediate future while having a broader outlook for the distant one. RHIP adapts a similar strategy, offering robust local planning and switching to determinate long-term planning. 🧠🔄 The outcome is a win for routing! Especially for sustainable transportation modes where factors like safety, road quality, and more play crucial roles. 🍃🚲 Case in point: Nottingham, UK. Previously, a preferred route was marked as private due to a gate. However, real-world data showed users frequently use this route without issues. Thanks to IRL, Google Maps now recognizes this and suggests the more efficient route. 🇬🇧🚦 In conclusion, scalability has always been a challenge in machine learning. Yet, with these advancements, Google Maps now trains on problems with hundreds of millions of parameters. To the tech community, this represents an unparalleled leap in IRL application for real-world settings! 🌐🔝 For a deep dive into this innovative journey, check out their paper. Kudos to the team for redefining routing! 🎉📊 https://lnkd.in/eCfyqpdW #MachineLearning #IRL #RoutingRevolution #TechInnovation

📱✨ In the digital realm, smartphones and browsers are our windows to information. Yet, navigating through a sea of clutter on websites can hamper our journey. This challenge amplifies especially for those needing better accessibility. Enter the advanced Reading Mode for #Android and #Chrome! 🚀 By leveraging Reading Mode, users enjoy better content clarity, customizable text sizes, legible fonts, and the power of text-to-speech. What's game-changing? Its ability to transform content on the go, without any data being sent out from the device. 📖🔄 Recently, a breakthrough has been achieved! Instead of the traditional DOM Distiller approach, a cutting-edge on-device content distillation model has been introduced. This new model taps into the potential of #GraphNeuralNetworks (GNNs) for content distillation. 🧠🔗 GNNs excel in understanding intricate relationships within data structures like trees. By deploying this tech on accessibility trees, a more streamlined representation of content is achieved. As a result, readers get a clean, distilled version of articles, without missing out on essential information. 🌳➡️📄 A visual showcase: 1️⃣ Extract tree representation of an article. 2️⃣ Compute lightweight node features. 3️⃣ Let the message-passing neural network navigate through the tree. 4️⃣ Classify content nodes as essential or not. 5️⃣ Present a decluttered version based on GNN insights. 📱💡 📈 Results? For Android, an F1-score exceeding 0.9 means 88% of articles are processed flawlessly. Comparisons with other models like DOM Distiller reveal the superior quality of this new approach. Even better, it supports a multitude of languages! In conclusion, as the digital horizon expands, tools like this underline the critical balance between user experience and data privacy. As we continue to ride the waves of digital content, prioritizing user needs remains at the forefront. Kudos to the team pushing these boundaries! 🌍🔒🎉 https://lnkd.in/gA_A35R8 #DigitalTransformation #ReadingMode #ContentDistillation #Accessibility #Innovation #DataPrivacy #Google #LLMs #MachineLearning #llms #Google

🔎 Exploring Algorithmic Reasoning in Large Language Models 🧠 The arena of Large Language Models (LLMs) like GPT-3 and PaLM is a buzzing space of progress and potential. One pervasive question is their ability to reason algorithmically, going beyond pattern recognition and honing in on rule-based comprehension. Can these models truly understand and apply the abstract rules defining an algorithm? 🌏 In a groundbreaking study titled “Teaching Algorithmic Reasoning via In-Context Learning”, researchers introduce an unprecedented approach leveraging in-context learning to imbue LLMs with algorithmic reasoning skills. The new strategy is focused on #InContextLearning - a method where the LLM learns to perform a task through understanding a few instances within its own context, without necessitating weight updates. 🎯 Delving into Algorithmic Prompting Building upon rationale-augmented methods, the team has pioneered algorithmic prompting, a technique that commands two distinctive features: - It furnishes detailed, step-by-step algorithmic solutions to tasks, and - It elucidates every step to prevent LLM misinterpretations. To illustrate, in two-number addition tasks, this prompting style not only processes each digit systematically but clarifies every move with explicit equations, ensuring no room for errors. 📊 Evidence of Enhanced Performance The experimentation has borne fruit, showing that LLMs can indeed answer correctly even with more challenging arithmetic problems when guided with algorithmic prompts. Not just that, they showcased adeptness at simulating multiplication algorithms by orchestrating a series of addition calculations. 🛠 Harnessing Algorithmic Skills for Complex Tasks Looking at a broader canvas, the researchers employed models specialized with different prompts interacting with one another to decipher complex tasks, spotlighting a triumphant strategy in addressing GSM8k-Hard math problems. This synergy facilitated a performance that was 2.3 times better than the existing baseline. ✅ A Bright Horizon This initiative paints a promising landscape for the future of LLMs, nudging them closer to executing tasks through input-agnostic algorithms and offering deeper, more intricate reasoning by embracing long contexts and providing richer explanations. https://lnkd.in/deD2cthE #LLMs #GPT3 #PaLM #AlgorithmicReasoning #InnovationInTechnology #BreakthroughResearch #FutureOfAI #NaturalLanguageProcessing #MachineLearning #InContextLearning Discover the full potential of large language models in advancing algorithmic reasoning in the complete study. Do share your thoughts and let’s #JoinTheConversation on the future pathways of #AI development.