There is quite a lot of Artificial Intelligence and Machine Learning being used in the actual update of the map. You need to essentially read through a stream of new inputs when you detect a change and figure out how to get the right update, right place and right time, says Stanimira Koleva, Senior Vice President (SVP) and General Manager Asia Pacific at HERE Technologies
HERE is one of the pioneers in embracing of Artificial Intelligence in the field of Location Intelligence. Can you shed some light on how you are using it and how it is empowering your solutions?
I think Big Data presents a big technology challenge. Take a step back and think of the up-to-date and high-quality maps for 200 countries and territories covering over 57 million kilometers created by HERE. Now, if you think about how many updates we need today and the terabytes of data that gets collected through our ongoing capturing, there is a massive technology problem in terms of how you not only record it and store it, but how you update it and publish services that sit on top of it. Today, other than AI, there are not too many viable technology solutions to crunch this amount of data. Once AI is applied to a massive problem that has many instances and challenges that you need algorithms to resolve, the algorithm gets trained and the capability improves.
The step we took to centralize a lot of this core content management and operations allowed us to be able to innovate at scale using Artificial Intelligence and Machine Learning, which also plays a major role. We have to deal with various inputs of data — video files, sensor data, crowdsourced data and more. So, there are many data types that we need to accommodate.
I am actually quite proud that a majority of such work is done in India, in Mumbai, which houses HERE’s Global Operations Center.
How big is the Mumbai center, and how do you see this growing in next five years?
We have about 4,500 people there and additionally we have access to another 1,600 outsourced workers. So, it is a fairly large operation for us. It is the heartbeat of how we do maps. We are one of the top 20 customers for AWS Storage, so you can imagine the amount of data being crunched there.
Mumbai is among our top three locations worldwide. While I do not have specific projection just now, we can safely say the operation will only grow in terms of capacity, and not go leaner. I hope that we automate and use more cool technologies and do things in a smarter and better manner.
How is HERE’s data and expertise helping in the construction of some of these realistic AI models?
I think there is no way to not admit that operational knowledge and capabilities that a 30-year business accumulates over time act as a massive opportunity and base for automation. What’s important is the way of doing things — more efficiently with streamlined processes, getting things done in a better way with better quality. That has been the base. The move to globalize operations gave us exposure to different countries. As you know, the map is local and so we need to account for local specifics of addressing and landscape. I think that is quite a variety of problems which created an opportunity for us. Traditionally, we have been improving our core processes. In the last few years, we have seen rapid automation around it then application of AI to resolve operational as well as data services challenges.
Can you give some examples of a few of these products, if these can be termed products?
I think there is quite a lot of AI and ML being used in the actual update of the map. You need to essentially read through a stream of new inputs when you detect a change and figure out how to get the right update, right place and right time. I think it is being driven by ML to a great extent because you do not want to necessarily repeat too many of these processes that get standardized. On the AI front, there are a lot of new products being built. For example, if you have a 2D map or a 2D picture of a place, then we have the capabilities to reveal the 3D model of the same place. That is very cool because they essentially have trained the algorithms to recognize from the shadow of the building how tall the structure is and recreate objects from the real world.
There are many products around predictability in travel. Taking into account multiple factors — your start the date, the day of the week, time of the day; it is the next version of the estimated arrival time, but it is much more complex and much more definite to give a precise estimate. There were some innovations that were coming in working backwards. It gives you suggestive departure time as well and suggests routing. We have a long list of challenges that Smart Cities are facing. For example, dealing with pollution. We are already working with Smart Cities essentially on how to give them re-route suggestions, particularly traffic re-route suggestions to decrease pollution hotspots and shorten the travel time.
From fleet operations, supply chain, automation to urban mobility, where exactly have your AI-powered operations proved most effective?
There are many solutions that are existent and are good. One thing that we are trying to resolve at the moment based on historic and static data — and we plan to move to real time data — is multi-model transportation services. We have found a lot of interest coming from operators, who have to, for instance, get their cargo on a truck that goes to a port and gets on to a ship and maybe sometimes s on a plane. The use cases concerning the prediction of timing around that, especially across countries and across the world, are gaining popularity. And we are developing that.
It is the same within the city in terms of private and public transport models. These are the use cases that we are seeing at the moment. I often also find a lot of interest around capabilities that we can project outdoor and indoor. Especially when you talk about new use case like food delivery and how you find a place and optimize the whole journey because most of these products are up to the gate or door. Last mile optimization is becoming very important and new technologies also emerging in this field. Especially in a country like India, if you come up with ways of dynamic generating addressing references for delivery purposes, positioning or location purposes, it would be fabulous. It would of course be an AI-driven capability because you do not deal with static datasets.
Can you tell us about the challenges that you have been facing in using AI?
One common challenge in using AI is the availability of enough data — the streams of data that you can use to train the algorithm. That is why we started with core map operations where we already had terabytes of data. With most of the new services, you need to take them to a level of maturity before you can apply some extra capabilities driven by AI, once you have underline stream of data. We now talk about data scientists. It is a fairly new skill that has been ramping up. It is only in the last few years that we see the availability of data scientist skills at any reasonable scale. We have around 70-75 data scientists at the Mumbai center. The availability of the deep skillset around AI is a limiting factor.
And then outside of it, you also have to give them enough opportunities to work. Not only from giving them the basics around compute power, storage and infrastructure capabilities, but you also have to give them meaty enough problems to resolve. In the last few years, this process of innovation has been improving and our team has organized itself into small groups of five-ten people of different skillsets. They do structured way of innovation days, innovation fares, patent filings and competitions. There are lot of activities that are targeted towards people giving their best.
When it comes to adoption of AI-based Location Intelligence, how would you rate Asia-Pacific as compared to the developed world, and how different it is to work with the private sector and governments in this region?
It gets down to solving different problems in different regions. For example, in Europe, we would have three-four use cases concerning pollution. In Asia-Pacific, we have use cases concerning the last mile for e-commerce and improving navigation for two-wheelers. All of these consume AI-based capabilities, but they resolve different problems. Even with standard problems like Smart Cities, they start with different priorities.
In India, we have been extensively talking to the National Informatics Centre (NIC). They have a massive database, but it is also not optimized in terms of what are the top issues. Some of it is even in real estate — positioning and verifications in order to support property registrations and transactions. They want to serve their agencies for different purposes. So, we are looking to partner with someone like them so that we can offer them capabilities. If you have a massive database, you will want to try to bring some capabilities from top of the data to give services to agencies. I see that is probably for them building the fundamental layer, but in many countries, it is already done. Digitalization efforts within the government also is a function of when many of these key efforts started. I would not say that the difference is necessarily technology, it is about just different use cases that you need to resolve first.