Category: Machine Learning

Source: Google

We live in a golden age for research.

Never before have we had the opportunity to make such rapid advances in computer science, and apply them so quickly to global-scale challenges, in a way that can help people in their daily lives. Since the start of my career, I’ve been excited by the “magic cycle” of research, where real-world challenges motivate new foundational and applied research, which in turn has a positive impact in the real world. Today, with the right infrastructure, talent and approach, we’re able not only to make rapid breakthroughs in everything from AI to algorithms to computing infrastructure, but to put those technologies to work to improve people’s daily lives and have societal impact faster than ever before, sometimes in a matter of months.

I’m seeing this firsthand as I’ve recently stepped up to lead Google Research, so I wanted to share a perspective on the incredible progress we’re seeing — and how important research is in driving helpful innovation.

Our approach: impact-driven curiosity

Google itself in fact began with research. “The anatomy of a large-scale hypertextual Web search engine,” published in 1998, explored how PageRank could provide a fundamentally better way to find info on the web, But it didn’t stop with a research paper — it was applying that research that produced Google. Over the past 26 years, that approach to implementing research — quickly — has transformed not only our company, but also how people can interact with the world’s information. Indeed, much of the rapid progress in AI we see all around us today grew from Google Research’s invention of the Transformer.

In all of our research, we ask ourselves: How can we make a step change, not just incremental? What’s impossible today, that we could make possible? And what is the greatest impact we can have — how can this make a real difference in the world?

Google Research today includes fundamental and applied work in foundational machine learning and algorithms, computing systems and quantum AI, and science, AI & societal impact. And across all these domains, we run initiatives on efficiency in machine learning, factuality & grounding in AI systems, and new data techniques.

Breakthroughs for the benefit of people and the planet

We motivate our research by going after the biggest questions that matter to advance science and make a difference to people and to communities globally. What are the most effective ways to mitigate climate change? How can we help make billions of people healthier? How can we enable new experiences? And to do all this, can we break through limitations in computing and AI systems? Each of those becomes an inspiring research challenge — and in so many cases we’ve already translated research into solutions.

For example, to address climate change, in a trial with American Airlines we used AI to help reduce contrails by 54%, demonstrating that airlines can verifiably avoid contrails and thereby reduce their climate impact. To help address the growing wildfire crisis, we partnered with leading wildfire authorities to develop FireSat, an upcoming AI-powered global satellite constellation to detect and track wildfires the size of a classroom — within 20 minutes. And consider flood forecasting — when our team at Google Research began the project in 2018, experts I met with said it was likely impossible to forecast riverine floods. But by developing AI that can build a global hydrologic model, we’ve not only proven it’s possible, but applied it to provide communities accurate early warnings and help save lives.

Meanwhile, to support health and wellbeing, we’ve developed AI that can help screen for breast cancer and colorectal cancer, help prevent blindness at scale, spot potential skin conditions and detect diseases based on the sound of coughs. We’re still in the earliest days of AI breakthroughs and genomics research, and yet we’ve made progress with Large Language Models for the medical domain and we’re already poised to improve the health care for hundreds of millions of people.

But perhaps one of the biggest undertakings involves advancing computing itself, and how it can fundamentally alter the scope of problem-solving. We’ve developed state-of-the-art attention models and use graph mining to improve retrieval quality. We’re also working on approaches to grounding large language models, such as by training models to rely on source documents for summarization and evaluating factual consistency. This research has led to improvements like the double-check feature in the Gemini app. We’ve made AI more efficient with research on speculative decoding, efficient inference techniques, and cascades, and we’ve helped improve privacy with federated learning and differential privacy. And our quantum computing team just published new results in Nature affirming that as we reduce the error rate in our quantum processors, we reach levels of computation that can’t be matched by even the world’s most powerful classical computers — putting us on track to crack open an entirely new computational capability for the world.

These are just a few examples of the type of work done across Google Research.

Why partnership is crucial for turning research into impact

Of course, as we turn research towards impact, we’re acutely aware that technology is not automatically beneficial — you can’t “invent it and forget it,” simply releasing powerful technologies on the naive assumption that they will be helpful. It requires careful stewardship, partnership and a fundamentally human-centric view of how to design and guide emerging technologies. That’s one reason we do our research in partnership with a multitude of researchers in academia and many others, while creating tools and best practices that support a truly global research ecosystem with the aim of steering new technologies towards beneficial outcomes. We actively engage in advancing socio-technical research to bolster our AI principles — including on equitable datasets, interpretability, and algorithmic fairness — and there’s important work ahead to make our AI models even more efficient, factual, robust and safe.

We have the greatest impact when we’re working with research partners. Since 2005, Google has worked with more than 1,000 research institutions and invested over $400 million dollars in academic research worldwide, much of this led by Google Research. We find experts across disciplines, roll up our sleeves, dive into the research, and make scientific advances together. In our connectomics research, we’ve partnered with Harvard to use AI to make the most detailed mapping yet of neurons in the human brain, revealing newly discovered structures — all towards helping scientists understand fundamental processes such as thought, learning and memory. Google Research also engaged in a first-of-its-kind collaboration with Howard University and other HBCUs to build a high-quality African-American English (AAE) speech dataset that Google — and others — will use to improve speech products. This is related to our overall effort to reduce barriers and better serve communities by enabling technology to work in many more languages.

With our partners, and through Google’s own products, we use our research advances to benefit billions of people. For example, as populations swell and shift in the Global South, millions of people’s buildings aren’t represented on any map, and they risk missing out on essentials like electricity, healthcare and mail delivery. So Google Research in Africa has used AI to make big improvements to the Open Buildings dataset — transforming blurry, low-res satellite imagery into useful, high-res building outlines so partners like the World Bank, the World Resources Institute, UN Habitat, WorldPop and Sunbird AI can use it to ensure global development includes everyone. Along with our SKAI effort, this has helped our partnership with the UN to boost damage assessment.

In another sphere, our Operations Research team recently showed how cargo shippers could double their profit, deliver 13% more containers and use 15% fewer ships. This is not only helpful for businesses, but good for supply chains globally.

Finally, we of course partner extensively with product teams to drive innovation across Google — and our responsibility also includes keeping an eye on the horizon, exploring the art of the possible, and imagining how we can apply our breakthrough technologies for maximum benefit in years to come.

Towards the future

We feel great urgency given the scope of problems facing humanity — but also great optimism because of what we’ve been able to do already. Our multi-decade track record shows that Google Research is second-to-none in delivering helpful advances. We’ve delivered breakthroughs that have shaped Google’s identity as a company, helped inaugurate new fields in computer science, and advanced the frontiers of innovation and technology with thousands of publications. The advances we’ve shared are already assisting people – from their everyday tasks, to their most ambitious and imaginative endeavors — and addressing society’s most pressing challenges, from healthcare to education to climate change and climate science.

We’ll keep sharing our breakthroughs on our Google Research blog, at conferences and at other events. We’re eager to explore — and invent — the future alongside all the partners and communities we work with.