Sella Nevo, a software engineer who specializes in machine learning (ML) research and development, currently leads the Google Flood Forecasting Initiative that aims to provide flood forecasts and warnings in developing countries. He was also one of the co-creators of the ML model used in Google Duplex. In a phone interview from Israel, Nevo--a keynote speaker at the Mint Digital Innovation Summit explains why Google chose Patna in India as a pilot for his project, and how the company plans to scale up this model not only in India but globally too. Edited excerpts:
What’s the goal of Google’s Flood Forecasting project?
This Flood Forecasting Initiative is Google's effort to provide high accuracy, high resolution flood forecasting. It's not global yet, and our focus is in using the Google's machine learning (ML) expertise and our computational power as well as our access to various types of resources and data to substantially improve flood forecasting systems, their accuracy, their lead time and so on.
Why did you choose India, and specifically Patna, to start this pilot?
We decided to start in India because the effect of flooding in this country is immense--about 20% of fatalities worldwide from flooding occur in India. We also believe that India is an incredible place for support innovation--it's a place where both the usage of the internet access is increasing incredibly fast and there is a lot of room for both collaborating with the government for innovative solutions and being able to provide value to individuals on the ground.
We chose Patna in India for several reasons. One is we wanted to start with a location where the rain is likely to cause flooding so that we know we'll be able to provide assistance in the short term as we prepare to scale up in the long term. We also found that Patna has an incredibly interesting and challenging location in the sense that there are a lot of things like embankments and other man-made structures that we need to be able to deal with.
What's the progress you have made so far? How are you planning to scale up this pilot?
The progress we've made so far is mainly exemplified by the pilot that we've done in Patna--there was one significant event that the Central Water Commission sent an alert for this past monsoon season in September. For that event, we sent an alert to people within about 1,000 square kilometers around Patna and the alert had a map that indicated which areas are likely to get flooded or are somewhat likely to get flooded and which areas are not likely to get flooded.
We're very happy with the accuracy of our modeling--broadly it was over 90% accurate across different metrics that we use to measure our accuracy. Our main focus now with the system is to scale it up to additional locations. Our goal is to launch in as many places as we can within the Ganges and Brahmaputra basins.
What kind of help are you getting from the Indian government in this regard?
Our collaboration with the Indian government includes two things. One is they share with us the data that is also accessible online. We just collaborate to make sure that the data transfer is reliable but they share with us the both the stream gauge measurements as well as forecasts based on their gauge-to-gauge forecasting system. We use these as inputs to our system which then outputs the more especially accurate map to know which areas exactly are going to be inundated. The second thing is that it is the mandate of the Indian government, and specifically that of the Central Water Commission, to provide information to people through a system. We see this project as empowering them (the government) in doing that.
There are quite a number of new technologies and methods for developing flood forecast maps. What's the Google approach, and how unique is it?
First is a new approach towards generating high resolution elevation maps. Google has developed a method of generating elevation maps at 1 meter resolution, based solely on completely standard optical imagery and that allows you to do this anywhere in the world. We currently try to update it annually where necessary, which is incredibly critical because a lot of rivers, especially ones with severe flooding, tend to be dynamic and change from year to year. The second area is on how to actually do what's called the hydraulic modeling more efficiently so that we can scale it up. Hydraulic modeling is the modeling of how the water will behave when it moves across the floodplain, which areas it will go to, and which areas are going to be affected and which ones are going to be safe.
What are the challenges and how are you using ML to address them?
Other than the challenge of getting elevation maps, there is also the challenge of computational complexity. If you want to have a high resolution model, it becomes incredibly computationally expensive.
We are working on using machine learning-based methods to make this modeling substantially more efficient than the classic finite element solution methods for solving this. We are also trying to do some remote discharge estimation, which is trying to estimate how much water has gone through a river based on satellite data and there we are using machine learning to integrate data from optical imagery, infrared imagery, radar, as well as microwave signals. So that is again a place where we think we are doing something that is totally different from that have been tried so far in this field.
Is this project basically the premise for showcasing the use of AI for social good?
I wouldn't say that its first goal is to showcase that. I think the first goal is to actually achieve social good in the sense of preventing fatalities and other harm that are currently caused by floods, but I think it is definitely an excellent example of how AI can be very, very impactful in these kinds of areas.
While we see, on the one hand, that AI is helping individuals, businesses and governments, there is simultaneously a lot of apprehension about AI being misused--the bias in algorithms and the impact of AI and robotics on jobs, the fear of a sentient AI overpowering humans, being cases in point. I would love your thoughts on this dystopian way of looking at AI.
I think sometimes AI can be an umbrella term to a large number of algorithms, topics and social issues and I think it might be useful to separate those out. I think there are a lot of very complicated issues--the interaction between AI and society. There're a lot of people who are much smarter than me and have a lot more to say about that, (and they) already disagree about what directions we’re going out towards as a society.
I don't know if I have something incredibly intelligent to say about that but I do think there are a lot of cases and I think that the parts that I know more of--and the things that I do day-to-day that are more related to these types of machine learning efforts--there are very, very clear and obvious things that we can do and we should do and which can have an incredibly positive impact. I think that if you don't put it under the kind of broad AI term umbrella, it's pretty clear to see that these do not create those same risks. I don't think that in these cases (such as flood forecasting), there is a genuine fear of robot uprising. (laughs).