Cloud weather: 3.28 million Yuan the stories behind the awards

For most people, weather and air quality monitoring has become a prerequisite software installed. Weather report false positives, misstatement of cases people also have become commonplace, many times, is not surprising. Despite the high tolerance of public weather forecast accuracy, but set-Chi Club in Beijing love AI technology geek, weather forecast accuracy still has a lot of room for improvement. Thus, Yuan Hang far in the Organization, set intellectual geeks have developed an application called cloud weather.
Now most of the weather application, is to show the weather in the coming days, cloud conditions through which the Central Meteorological Observatory radar image data analysis, precise weather forecasting is accurate to the minute trends, users can also see the recent rains across the city from precipitation map dynamically. Cloud weather interface is simple, the top is a mobile positioning location, is in Central time, weather and temperature, as well as “36 minutes after the rain stops” sort of prediction, then pull it down is the rainfall maps, dynamic display of rainfall for the past two hours.

By virtue of such a product, cloud weather team recently won the “2014 High German LBS application contest” 3.28 million dollars in awards. Recently, the CSDN interviewed cloud weather team, please talk to them about a weather the story behind the development of clouds, geeks and technology from the perspective of computer science and deep learning.
CSDN: in the United States had a very successful short-term weather forecast products Dark Sky, clouds whether the weather was born inspired by this product?
Yuan Hangyuan: development of high accuracy rain forecast was my original ideas, do not at first aware of the existence of darksky. Because I had always been more like weather, for example in the rainy season when I will pass the analysis of radar to judge small to my friends when rain so I thought it was valuable. In the past, when weather forecast meteorological stations on different place (such as the venues of the Olympic Games), you need to find forecaster on duty, costs tend to be higher, but if I let the machine go through radar data analysis of weather conditions at different locations, you will greatly reduce costs, bringing advanced technology to ordinary people. The other hand, I think cloud weather most the interesting part is, I don’t know if it can not be achieved in the end. For example, a business class or social App, regardless of the success of the business model, at least technically, this application can be implemented according to my idea. But this case is just the opposite of the product, I said want to forecast rainfall per minute, most people said was impossible, I don’t know whether it is technically able to achieve. This is what I would like to share with other developers, and that is to do something you do not know if I could do something, face the challenges, rather than stick has become a routine thing. (The legendary hard mode … … ）
About a month later have been made, I’ve seen Dark Sky,Dark Sky made me realize two things: 1. This thing is made out of 2. This stuff is used. Even so, I know for the first time after the Dark Sky, was a bit frustrated, and even wanted to give up because someone else had already made them. But on second thoughts I think domestic users are still unable to enjoy the services of Dark Sky, so knew of darksky more strengthened my determination to continue.
CSDN: do you mind being called China’s Dark Sky?
Yuan Hangyuan: not at all. English is a Colorful cloud weather Clouds,Dark Sky translates as “dark sky”. We deliberately keep it forms an “antithesis”, is also a kind of salute. Initially we name local area, and later felt too, was changed to cloud weather. In the preface of this name also comes from the poetic, “rain cloud pin JI, colour penetration area out” of the poem.
CSDN: tell me about your personal background.
Yuan Hangyuan: I was born in the Sichuan province, graduated from the Department of mathematics of Shandong University of science and technology, in Qingdao, after graduating from a local live search companies using screen work, mainly responsible for the search and data mining-related work, with a team of a dozen people. Then I thought Internet atmosphere in Beijing, Qingdao, and so came to Beijing from Qingdao, who works for Beijing Branch of Taobao. I search in Taobao’s work on commodities recommended, of course, also doing some dirty work, such as distributed computing platforms. In February 2013, I resigned from Taobao, would like to play a game, then went to Guangzhou, Shenzhen, Shanghai and Suzhou, among other places, because I love technology, so I went to places like the Daya Bay nuclear power plant, Minsk aircraft carrier. These two years have been volunteers LEAD Sun clubs in Beijing over the weekend will go to the surrounding Beijing migrant school interested in coaching, resigned later to LEAD the Sun to develop a collaborative lesson plan writing and sharing site “encyclopedia of sunshine.” Gapyear also learned some theory related to computational neuroscience, JI Zhi Club participated in a brain and deep learning, reading, cloud weather now part of team members used to be the core members of the book club.
CSDN: Central Meteorological Observatory has specialized data, almost all the weather App is using the same data source, so, how do you ensure accuracy of cloud weather predictions?
Yuan Hangyuan: indeed, most of the weather data used by such applications are forecast results from the met Office, but there are some enthusiasts will go to see something called “weather radar base reflectivity” data. It is also open to the public, but did not disclose the latitude and longitude, and RADIUS, so more difficult to deal with. On our interactive map search radar, and get coordinates with the radar control, made with machine-processed radar, this higher-precision of possible data sources. But our cooperation in September 2014 has been reached and the China Meteorological Administration, minutes are currently has started collaborating with the met Office forecasting technology research and development, data sources, and forecast a further improvement in the accuracy of the results should be.
CSDN: this technological threshold is very high.
Yuan Hangyuan: Yes, just find the data source, it can hold up to 90% people. It’s like writing articles to find pen and paper before you, then a lot of people say, just get the radar chart, remaining minutes can do the job, but this is just a pen and paper. Find pen and paper only to find many people to write articles, how to write the article? The technical difficulty compared to find pen and paper is not an order of magnitude. Said positioning, noise reduction and forecasts of rain after a positioning belt movement. We will send a reminder to the user when the rain, such as telling the user “after 30 minutes of rain”, then we will put a time of 40 minutes, if it rains during this period, describes the prediction right. Yesterday, we sent more than 1500 pieces of reminders, 96% is right.
Data only as raw materials, predicted results of the dish. Below is an example shows one June day Yahoo and the Central Weather Bureau forecast results. There was a sudden storm, but paused a while and can clearly see that they were directed at the entire city can be predicted and did not take into account the heavy rains suddenly.

CSDN: so it seems that weather the clouds far outweigh the value of B2C area.
Yuan Hangyuan: so to speak. In July of this year, the Central Weather Bureau’s Center for new media in touch with us, we hope that both sides can work together. This cooperation has been reached, now the public can see on the weather network, owned by the Central Weather Bureau the results of our cooperation, or open the Central Weather Bureau’s official App inside minute levels of high-precision weather forecasting is ours, and jointly launched by the met Office. Our goals while maintaining high accuracy, forecast extended to 2-3 hours, another forecast of hail, snow and other extreme environments. The future, we also hope to give users related to the provision of meteorological early warning weather data services.
CSDN: cloud weather using Convolutional neural networks, in your opinion, what deep learning is the most interesting?
Yuan Hangyuan: I think the machine is the learning ability. You throw it two pictures, not tell it was the wind, what is the wind direction, what is the terrain, what is the air pressure, it is only through the statistics of rainfall patterns in nature, it will tell you the probability of rain. As if we are doing a math problem, you need to learn Mathematics, and write formulas, have been extrapolated to the results. But one day, you just tell the computer topics and results, it saw a heap of problems later, you give it a title of it haven’t seen, it can also know what the answer is. Convolutional neural networks we use, if suddenly a cloud in the Beijing area, we can capture this phenomenon. This is the specific application, but the depth itself is not magic.
CSDN: why it is not magic?
Yuan Hangyuan: depth of learning to learn is patterns, using patterns to infer future phenomenon is not a new idea. But the development of the neural network algorithm is a process of quantitative change to qualitative change. In fact, we have been using multilayer neural network too, but with the advancement of the Internet and computing power and the amount of data has improved, so the sudden depth of learning will be a qualitative change. I think it’s a trend in the computer industry: computer science is computer algorithm computer algorithms of machine learning; learning machine learning is equal to depth. Are all down, cause when it comes to computer science, everyone thought that deep learning.
From its origin, Turing invented the tape that year model is actually the character’s writing, in can be represented on a piece of paper out of the thing, on a piece of paper you can get set up, if the tape is long enough, it can be said enough information. He initially did not need to build a computer, and his goal is a simulated calculation process, Turing brought him something abstract into a very pure, and then puts all the calculations and can write things, all can be done by this method so that the machine. But there is one significant disadvantage, that tape operation procedures must be written, that is, to have a rule. Such as computers with CPU,CPU there is a set of instructions, instruction set is fixed. Neural network development trend now is to make this fixed things are not fixed, as we discussed on Friday reading Club “neuron Turing”, with a neural networks training instruction set, this may be the depth of learning and development direction in the future.
CSDN: not long ago, you had just won the “2014 High German LBS application contest” 3.28 million dollars in awards, initially in cooperation with Golder what is the opportunity?
Yuan Hangyuan: over the years, I like gold. I features high dynamic routes to avoid congestion particularly impressed. This feature is nice, because the average route plan will only tell you the shortest path, traveling over the process of real-time road conditions change. On one occasion I was in Guangzhou, high open de navigation, it is suggested that a tunnel ahead, the driver was surprised, saying, “I don’t know in front of the tunnel.” Another time I walk on the four ring, gold prompted a path could not be found in the map, I rounded the last look, turns out to be forest fire prevention road, results going to all of a sudden. Examples like this very much. So when Google in no time, my first thought was to use gold to show cloud rainfall map and provide location capabilities. Compare with Google, Gaud SDK can save nearly 70% of traffic, this is a very big advantage. There are many clouds and clouds the weather app itself dynamic effects, require updates data every minute, this requires our applications perform better than similar products, and guide locate the SDK installation package is much lighter than similar products, loads very fast, and it helps to enhance the user experience of cloud weather. Cloud weather now in other parts of the world there are some services and users, and high-positioned before the SDK was the only company to provide global positioning services SDK, which coincides with cloud weather business development.
CSDN: in contact with gold, and what impresses you about?
Yuan Hangyuan: very early age, once a Bug we found gold, is next refreshed, maps in some degree. Then we don’t have a formal cooperation, so I wrote an email to guide this. But their response was very quick and committed to solve this Bug in the next version, and prior to that can help us to change the code, let the App up and running. In fact, gold can not solved, because we are not its main customer, but they solved this problem. After our formal cooperation, cloud version 1.3 there is a weather problem, touch gold, gold direct give me a competent business QQ, Liu Yewei lets me talk to him directly responsible for technology at that time was on leave, but I immediately write an email, ask a colleague to solve this problem as soon as possible. Both times left me a very deep impression, even with the two departments of a company does not necessarily have a high response speed and cooperation with Golder, it feels like everyone in the same company Office.