After understanding the basic operation process of Alibaba Cloud artificial intelligence robot small Ai, the author believes that the algorithm of small Ai is not only deep learning, but a higher-order algorithm above deep learning, and it is directed to Aliyun artificial intelligence scientists. Wanli verification, get the reply: Yes.
Where is the high order of the small Ai algorithm?
Although Alibaba Cloud did not disclose more information about the algorithm of the small Ai, only the small Ai is mainly based on the principles of neural network, social computing (social compuTIng), emotional perception, etc., good at insight into the nature and real-time prediction, and can understand human emotions. .
But the small Ai challenged the song king to predict three difficulties: First, this is a melee between the seven singers, not the two-player game between AlphaGo and Li Shishi; secondly, singing is emotional understanding plus art appreciation, different singing, changing voices, False sounds and other impermanence methods are also inconclusive, and live singing is improvisation; again, the end result is the result of the joint creation of Hunan Satellite TV program group, TV audience, 500 public judges, and 7 singers, which is full of great randomness.
A simple understanding, the deep learning algorithm for AlphaGo Go game, speech recognition, image recognition, etc., mainly solves the single-objective optimization problem, and the small Ai challenge is essentially a multi-objective optimization problem, which fundamentally Forcing "Alibaba Cloud to choose a higher order algorithm. Yan Wanli told reporters that the small Ai itself is a probability-based decision-making optimization process, which requires multiple sets of variable inputs in the decision-making process, some of which are optimized through deep learning.
Previously, the algorithm of Small Ai has been successfully used by Alibaba Cloud for the road condition prediction of the Zhejiang Provincial Communications Department, and the artificial intelligence algorithms that can solve urban management and macroeconomic problems belong to the multi-objective optimization algorithm. The West has long adopted multi-objective optimization algorithms in areas such as urban integrated management and macro-policy setting, including simulated annealing algorithms and genetic algorithms. Yu Wanli revealed that the small Ai did not use the off-the-shelf multi-objective optimization algorithm, but the algorithm system developed by itself.
What is the small Ai?
So, what can Ai achieve high-order algorithms? This must mention the father of the small Ai, Aliyun artificial intelligence scientist Wan Wanli.
At the age of 14, he was admitted to the junior class of the Chinese University of Science and Technology. He went to the United States to study for a master's degree in physics at the age of 19, and received his Ph.D. in statistics from the University of Chicago in 2004. He was a researcher at the IBM Watson Institute and Google. In 2013, he joined Alibaba Cloud in charge of artificial intelligence projects. Ai.
Wan Wanli said that his experience at the IBM Watson Institute has benefited him a lot. IBM first proposed the smart city strategy, and it was the first to see the future trend. In 2005, after IBM sold its PC hardware business to Lenovo, it began its own transformation. The transformation process, the most lacking is the massive data analysis capabilities, so it complements many related research projects, including massive data analysis, key information extraction, predictive modeling, machine learning, etc., and Wanli also used this to contact the first-line actual cases.
Later, when I arrived at Google, Wan Wanli was more specifically responsible for the optimization research of mobile-side advertising precision push, mainly to judge the massive data, which is actually very close to the problem that small Ai has to solve today. Mobile-end advertising precision push optimization is the core technology of Internet advertising. It needs to improve the accurate push of advertisements through big data analysis and machine learning, thus improving the click rate of advertisements.
The mobile terminal advertisement accurate optimization data analysis includes: judging the content of the user's current context page; judging the user's geographical location - pushing short advertisements while driving, pushing complex advertisements in the restaurant; and judging the user's preference for pushing advertisements. These are all real-time decision problems under multi-scenario variables, which are very similar to the scenario predicted by the small Ai.
Yan Wanli has long been engaged in machine learning theory research and application algorithm research and development, and has obtained many international patents in the fields of electroencephalogram (EEG) analysis, high-dimensional data mining, stochastic process theory, time series analysis, and network flow theory. His road traffic flow prediction study published in 2011 is one of the most cited papers in the field in the world in five years.
In 2013, a headhunter found Wanli, hoping to lobby him to join Alibaba Cloud. At that time, the headhunter said that there is such a company in China, and the total amount of data is more than Amazon, eBay, and Paypal. "This sentence can already explain everything. To do big data, you have to come to Ali."
Small Ai development process
Since 2012, Alibaba Cloud has developed a massive data processing calculation engine, later called "MaxCompute", which was called ODPS, which is a very important computing infrastructure for small Ai.
ODPS is the only big data processing platform of more than 30 divisions of Ali Group. In the 2015 World Sort Benchmark sorting competition, ODPS completed 100TB data sorting in 377 seconds, breaking the 1406 second record created by Apache Spark and creating 4 World record. Today, MaxCompute can process 100PB of data, equivalent to 100 million HD movies in 6 hours.
It is worth mentioning that ODPS's real-time computing system StreamSQL, later known as "StreamCompute" in the Alibaba Cloud Data Plus platform, can process terabytes of messages, PB-level data, and tens of millions of queries per second QPS. A real-time recommendation system for adjusting recommended products based on user real-time behavior data (browsing, closing, collecting, etc.).
Alibaba Cloud said that the learning speed of the small Ai is 10,000 times that of human beings. The human needs 100,000 hours to become an expert in a certain field. The small Ai only takes 10 hours. This is actually based on the actual platform of Alibaba Cloud big data analysis such as MaxCompute and StreamCompute. "Alibaba Cloud's big data analysis is a test of actual combat, which is different from other platforms." This platform has experienced tens of thousands of engineers in Ali. The actual combat, including the test of seven double eleven.
While developing computing platforms such as MaxCompute, Alibaba Cloud Artificial Intelligence team is also developing artificial intelligence algorithm systems such as deep learning, social network sentiment analysis, semantic analysis, and optimization algorithms. By 2015, Alibaba Cloud's artificial intelligence algorithm implemented a mature application in various business scenarios in Ali, and then was abstracted into a generic module and then grafted onto MaxCompute.
"So Alibaba Cloud's artificial intelligence module has a clear applicationable scenario, not a closed-door car. It's not that we have been smashing for 4 years, just to do it (challenging "I am a singer"). This is actually the entire business ecosystem in Ali. In, slowly grow up." Wan Wanli said.
Small Ai algorithm system
Alibaba Cloud began technical research and development and reserves in the field of artificial intelligence in 2012. Before participating in the Hunan Satellite TV "I am a singer" competition, Xiao Ai has accumulated a lot of practical experience, such as helping photovoltaic power plants estimate power generation capacity to reduce energy consumption, helping water conservancy supervision departments to predict reservoir water level to prevent disasters and help financial institutions to help customers. The staff answered the phone, helped Ali music predict music black horse and so on.
In addition to engineers and scientists, the small Ai team has a number of Ali music and professionals as coaches, and has learned millions of songs to enhance their musical taste and appreciation. Based on the Ali music database, the small Ai automatically learns the important features of the audio to form a multi-dimensional evaluation of the song, including pitch, energy, speech, fundamental frequency, etc., through the feature to train the small Ai between the audio and the popularity. Relevance thinking, but at present, the small Ai has not traversed foreign songs such as Korean.
So how does the small Ai predict the outcome of the game on the spot? Small Ai looks for variable factors that influence the outcome of the game from historical events and massive data, and trains a real-time dynamic model for prediction, including songs, singers, fans, live atmosphere, and netizen discussions. Each dimension extracts massive amounts through machine learning. feature. These features are static and dynamically change following the game and require real-time calculations on site.
Yan Wanli revealed that the singer in the eyes of Xiao Ai is a collection of countless labels. For example, Li Wei has female singers, 70s, Chinese, American, idol, sexy, Oscar, R&B, Soul, fan volume and so on. There are many factors influencing the judges' preference, and the small Ai needs to find every information variable of these influencing factors, including the genre singing, genre, genre, arranger, singer, dance, singer, and other on-site hot data. And a variety of cold data information off the field. This information is then superimposed by certain logic to form an overall decision-making mechanism.
To sum up, the small Ai is going to observe all the factors that may affect the voting results. The whole process is to "understand human preferences" and "investigate human thinking" in a full-space, continuous, and dynamic scenario.
The difference with Microsoft Xiaobing
As Microsoft Xiao Bing, who also focuses on "emotional algorithms," Wan Wanli said that Microsoft's Xiao Bing's question-and-answer dialogue with humans can easily be brought into context, and then can be modeled and analyzed using related models such as linguistics. The small Ai has to understand the singer's melee and music appreciation and other language beyond the language. This is the difference between the small Ai.
é—µ Wanli exemplifies the difficulty of the small Ai algorithm. For example, the last Sun Nan retired completely unexpectedly, and the small Ai could not anticipate such an event, which caused great trouble to the model training at that time. In the live game, what happens is possible, regardless of the outcome of the final small Ai forecast, is a successful attempt.
In addition, in terms of business model, the small Ai does not seem to have an easier commercialization route as Microsoft Xiao Bing, and now Microsoft is turning Xiao Bing into an artificial intelligence infrastructure for Microsoft products and services. Compared with the current prevailing deep learning algorithms, the small Ai multi-objective optimization algorithm seems to be difficult to see a clear commercialization prospect.
For the commercialization model, Wan Wanli said that Ali has a strategy: Happiness and Health, which is the happiness index and health index. Little Ai's musical and artistic appreciation is related to both the happiness index and the health index. It is already a huge technical achievement to achieve this. In addition, the ability of the small Ai can be easily generalized into the commercial field, and has been successfully applied in traffic management, energy management, weather forecasting, etc.
"I am a singer" in the fourth quarter of the finals of the small and medium Ai real-time forecasting attempts have been widely concerned, IBM also sent a wish for the small Ai in advance. More importantly, this prediction is a challenge for multi-objective optimization algorithms, and it can be said that it was also experienced by the public for the first time in China.
Since then, artificial intelligence has opened a new chapter.
new smart watch,sport smart watch,smart watch supplies,phone smart watch for Iphone
Shenzhen Konchang Electronic Technology Co.,Ltd , https://www.konchang.com