Text-to-Image Model, Language Translation and Other Highlights from Google’s AI Event

On Wednesday, Google hosted its first AI@ event to showcase its latest work in AI. The event focused on new developments as well as early researches from Google AI and research teams in areas of climate change, natural language understanding, accessibility, healthcare, and creativity. 

Additionally, the event focused heavily on Google’s efforts to build responsible AI for prioritising the safety and privacy of people amidst rapid advancements in AI. 

Google’s product innovation branched into three categories: AI for Social Good, AI for Creativity, and AI for Language Inclusivity. 

The full video of the event can be accessed here.

Key highlights

• Google announced Floodhub, a platform that brings flood alerts to the general public. Since 2018, Google has been involved in applying AI for flood forecasting, alerting users via Google search and Maps. But now, using new advances in AI, it plans to expand its coverage to 18 new countries. The new platform will use satellite imagery and AI to accurately map rivers, model what areas will be flooded, and send timely info. 

• Google’s satellite imagery technology will also train AI models to identify and track wildfires in real time, thus taking a foot forward in addressing the impact of climate change. The wildfire tracking system will be able to predict how the fire will evolve and spread, helping people and firefighters take timely actions against it. 

More information on the precise nature of AI models built by the Google researchers, and their work with fire authorities can be found here. 

• Leveraging AI to bring high-quality, equitable healthcare across the globe has also been a part of Google’s developmental agenda. In partnership with Northwestern Medicine, the company is developing a low-cost ultrasonic device aimed towards assisting nurses and midwives in regions without sonographers. The AI application will help healthcare workers collect and interpret ultrasound images, and provide timely care by identifying issues early on in pregnancy. 

Google also plans on expanding its reach in their continuous efforts towards increasing healthcare accessibility. It will increase partnership with caregivers and public health agencies to provide access to diabetic retinopathy screening through their Automated Retinal Disease Assessment tool (ARDA). 

Further, in its efforts to unlock healthcare access to anyone with a smartphone, Google is working on eye-screening, as well as, respiratory and heart-rate tracking tools. 

• Google also announced the Imagen text-to-image model, which differs from its earlier approach called Parti. Imagen uses diffusion models to create high resolution images mapping noise back to data. This new AI model will be coming soon to AI Test Kitchen, an app where users can learn, experience and give feedback on Google’s emerging AI technology. It also, for the first time, rendered a long-form coherent video from text prompts, using Imagen Video, and Phenaki, a language model that turns text into videos. 

• The event included the latest developments in AI research built on the Pathways Language Model (PaLM). Google’s Minerva AI system is based on PaLM and is used to solve complex mathematical problems using symbolic and arithmetic reasoning. Google is also experimenting with the application of PaLM in robotics. PaLM-SayCan is a robot learning model combining the natural language capabilities of PaLM with real-world capabilities of helper robots. 

• Going forward, Google is also planning to build a model that can support 1000 most-spoken languages through its 1000 Languages Initiative. The initiative will focus on increasing representation in training AI models.