AI Innovations Drive Climate Resilience, Genetic Discoveries, and Smarter Navigation

Google Research is leveraging AI to enhance climate resilience by improving predictions and mitigation strategies for climate-related threats. Their AI-powered flood forecasting model now covers over 700 million people across 100+ countries. They're also improving cyclone forecasts, aiming for earlier and more accurate warnings, and applying AI to short-term weather predictions, particularly in Africa. Furthermore, they're using AI and satellite imagery to detect wildfires, with the FireSat constellation set to provide high-resolution, near-real-time data, and geospatial reasoning to enable planetary insights. Finally, AI is also used to reduce emissions and improve air quality.

Google Research introduces MUVERA, a novel algorithm that significantly speeds up multi-vector retrieval by transforming it into a single-vector search problem. By constructing fixed dimensional encodings (FDEs) of queries and documents, MUVERA allows the use of optimized maximum inner product search (MIPS) algorithms. This approach achieves high retrieval accuracy with substantially reduced latency compared to state-of-the-art methods, outperforming existing techniques by up to 10% in recall and reducing latency by 90% across BEIR datasets. An open-source implementation is available.

Google Research introduces a novel generative AI method for efficient regional environmental risk assessment. By combining physics-based climate modeling with probabilistic diffusion models, the approach bridges the resolution gap between Earth system models and local needs. This dynamical-generative downscaling method produces detailed local environmental risk assessments at a fraction of the cost of existing techniques. The AI system, called R2D2, learns to add realistic, fine-scale details to intermediate-resolution output, efficiently bringing it up to the target high resolution. This enables better-informed decisions for adaptation and resilience policies across vital sectors.

M-REGLE, a multimodal AI method, enhances genetic discovery by simultaneously analyzing diverse health data streams like ECG and PPG. By jointly learning from multiple data types, M-REGLE creates richer representations, boosting the discovery of genetic links to diseases such as atrial fibrillation. It uses a CVAE to learn compressed signatures from combined data streams, reducing reconstruction errors and improving the identification of genetic associations compared to unimodal approaches. M-REGLE holds promise for predicting disease risk and identifying new therapeutic targets, especially with the rise of smart wearables.

Google Maps has introduced a new feature providing HOV-specific routing and ETAs, improving overall ETA accuracy for drivers using HOV lanes by 75%. This was achieved through a novel classification approach that infers HOV travel times by analyzing traffic trends and uses a mixture-of-experts model to classify trips based on speed, lateral distance, and temporal clustering. The new model addresses the scarcity of labeled HOV data and improves the accuracy of estimated arrival times by 18% compared to previous methods, contributing to smarter and greener commuting.