Scaling Language-Image Learning in 100 Languages with PaLI @GoogleResearch

Uploaded October 2022 | Updated September 2026, 2 weeks ago

Have you ever wondered how teams across Google Research worked together to develop PaLI, the Pathways Language and Image Model, as a scalable multilingual approach to joint language and vision modeling? In this episode of #ResearchBytes, Research Scientists Radu Soricut and Neil Houlsby explain how PaLI works, share insights from their research combining language and vision models, and more! Watch this episode to learn more about PaLI and how it was created!

Resources:

Read the PaLI blog post → https://goo.gle/3sKmrgl
Read the paper → https://goo.gle/3WeCmkL
Learn how it works →https://goo.gle/3SRTtpF

Chapters:

0:00 - Intro
0:36 - What was the challenge with PaLI?
3:04 - What did we do?
4:14 - How did we do it?
9:40 - What’s next?

Watch more episodes of ResearchBytes → https://goo.gle/ResearchBytes
Subscribe to the Google Research Channel → https://goo.gle/GoogleResearch

#ResearchBytes

Scaling Language-Image Learning in 100 Languages with PaLI

$Making Weather Forecasts More Accurate Climate change is real, and its effects are becoming increasingly evident. Traditional climate models struggle to produce clear, accurate pictures of how our climate will continue to change due to limitations in how the models represent complex phenomena. This video explores NeuralGCM, a groundbreaking AI-powered approach developed by Google Research that could someday offer a faster, more efficient, and more accurate way to predict climate change. Learn more about NeuralGCM in Nature! You’ll hear from: Stephan Hoyer, PhD, AI weather/climate lead at Google Research Janni Yuval, PhD, visiting Google Research Scientist Watch more Research Bytes → https://goo.gle/ResearchBytes Subscribe to the Google Research Channel → https://goo.gle/GoogleResearch The Challenge of Climate Prediction Accurate climate predictions are a matter of urgency because they could allow us to prepare and adapt to climate change. NeuralGCM tackles the limitations of current physics-based models by combining them with machine learning. This allows for faster simulations and more accurate results. How Does NeuralGCM Work? Traditional climate models divide the world into large cubes, simulating the climate in each and sometimes making approximations about small-scale phenomena like clouds. NeuralGCM uses AI to make better approximations by learning about these phenomena from past data patterns, ultimately leading to a more accurate prediction engine. A Game-Changer for Climate Science We’ll dive into the revolutionary potential of NeuralGCM. It is more accurate than traditional models, reduces bias, and achieves supercomputer-level accuracy while using a fraction of the computational power. This allows researchers to simulate thousands of days in the same amount of time it would take them to simulate just one day using a high-resolution physics model, significantly improving our understanding of future climate scenarios. Benefits for Humanity NeuralGCM is a groundbreaking step toward a new approach of building climate models that could make them faster and more accurate. Improving climate models would empower us to understand and prepare for the impacts of climate change. We can mitigate global warming by reducing carbon emissions and adapt to its effects through regional planning and infrastructure development. Open-source research like NeuralGCM supports Googles commitment to making this technology universally accessible. This video is perfect for anyone interested in: Climate change and its impact Artificial intelligence applications The future of weather prediction Technological solutions for a sustainable future #AI #climateaction #GoogleResearch #climatescience #machinelearning #AIforsocialgood #globalwarming #climatestrategy #ECMWF #naturemagazine$

How to get started with the Google FloodHub

Project Relate - Communication made easier

Open Buildings: Using AI to put everyone on the map

Quantum Resource Estimation: Paving the Way for Fusion Energy and Beyond

Ray Kurzweil on the exponential growth of computing power

How AI can help policymakers improve access to emergency obstetric care