“I am often asked, what is the future of AI? What are your expectations for the future? I usually say that the development of the financial technology companies in the ecological field will be the most interesting place for AI in the future.†Machine Learning Michael I. Jordan, a leading scholar in the field, chairman of the Ant Financial Science Think Tank, and a professor at the University of California, Berkeley, said.
Indeed, these companies face real industry challenges, and they are where AI applications are most effective. “We began to discover that AI is no longer simply trying to imitate human intelligence, but has become a new capability. It permeates every corner of the social economy like water, dissipating intelligence and solving problems.†Ant Gold Service CTO Cheng Li explained,
"Based on the ability released by such AI, it not only helps people to have a deeper insight into society, but also better discovers and controls risks in the financial field. It can also create an intimate life assistant and wealth consultant for every ordinary person. Every small and micro enterprise has the most intelligent business assistants and the best CFOs."
On April 25th, the ATEC Ant Developer Competition artificial intelligence contest was launched. The original intention of the organizer Ant Financial Service was to extract the most valuable problems in the business system and provide an autonomous and controllable AI platform for sharing with developers. Innovative technical solutions to solve.
"In fact, AI is an engineering discipline. Sometimes everyone thinks that AI is a magical tool. Bringing it to the company will bring miracles. In fact, this is not true." Michael I. Jordan said, "AI is engineered. The brain is a means of solving problems. It is a way to think about solving problems. Today's competition is a very good example. You can think about how to solve problems with engineers' minds instead of using magic. The means bring miracles."
To a certain extent, for these contestants, it is more like participating in a competition, more like completing a real project in a technology finance company. We got a deep understanding of the competition around the specific financial business scene, and found some solutions to the problem from the experts.
First, payment risk identificationThe first question is a question about the identification of payment risks.
While the new financial services based on the mobile Internet are booming, the ability to attack black products is also constantly upgrading. Information disclosure is one of the most serious issues that lead to account security. In the international incident of user information leakage, over 100 million sensitive data has been stolen and used, which has brought huge economic losses to users and banks.
Protecting consumers and identifying risks is increasingly becoming the focus of attention in the financial industry and academia. Starting from this research point, Ant Financial will highly simulate the data and scenes of real business. I hope that the industry will make in-depth research and breakthroughs on the two technical problems of risk mode and unlabeled data learning. The industry brings value while promoting the next step in the development of artificial intelligence.
“This has been a big problem for the past four decades, and it has not been solved yet. In the 1990s, some banks began to use machine learning for testing, and the results were good. After Amazon’s online use, we can see false payments. The proportion is increasing, they use machine learning to reduce the false payment rate a lot," said Michael I. Jordan.
At the event site, Yan Xiongwen, vice president of ant Jinfu and head of security, introduced the background of the competition.
Alipay is the origin of Ant Financial, and the fundamental reason why mobile payment is popular in China is that compared to credit cards, mobile payment based on QR code has lowered the payment threshold for consumers and merchants, and compared with POS. It costs less.
On the basis of convenience and low cost, safety has become the primary problem faced by Ant Financial, which is the cornerstone of all financial services.
In terms of security technology, Ant Financial has three core competencies: ensuring the digital identity of the operator in the online and offline environment, ie the account owner; using big data, machine learning and artificial intelligence to achieve accurate identification and risk prevention. Intelligent wind-controlled brain; and data protection and privacy from internal and external attacks.
Yan Xiongwen also mentioned the latest exploration of ant Jinfu in intelligent wind control - AlphaRisk wind control engine.
The engine has four core systems, including risk perception (perceived risk situation and situation), risk identification (mm-level accurate identification of all risks), and autonomous driving (management measures based on risk scenarios and user status to complete personalized smart recommendations). And self-learning ability (to ensure that the entire wind control system can automatically learn and update while the black production changes constantly).
With the support of this set of wind control engines, the current ALS service has a loss rate of less than one in a million. It can process more than 500,000 risk transactions per second at the peak, and can replace the entire strategy within 1 second. system.
However, there are still some difficult challenges in the field of intelligent risk control. Yan Xiongwen cited two examples. One is that the means of black production is constantly changing. The traditional offensive and defensive model will decrease in a few days. Second, the model will be accurately identified in the absence of black samples.
In this regard, Jiang Changjun, the president of Donghua University and former vice president of Tongji University, also shared his views:
"The rules system used in the early days is product-oriented and targeted, but adaptability and evolution are weak. However, it is impossible for a scammer to take a trick to get rid of it once and for all, and it will definitely be updated. At this time, the rule system is difficult to adapt, so models and algorithms are needed. solve this problem."
However, the model and algorithm are extracted from a mathematical processing method, and its scene is not so strong. In this process, how to integrate the rules and scene features into the model and algorithm, and adapt the model and algorithm to the whole scene, will improve the timeliness and stability of the system.
"This is the place where the drivers need to be smart." Jiang Changjun said,
“The other is the problem of a serious asymmetry between the black sample and the normal sample size. In general, the model built by the two samples under similar conditions will be better, and in the case of severe asymmetry, how can this effect be achieved? This is the second difficulty of this risk control problem. It is also the highlight of this problem. I think this is a place on the issue of intelligent risk control that requires everyone to innovate and create."
Second, intelligent financial customer serviceThe second game is related to intelligent customer service. The important goal of Inclusive Finance is to provide efficient and personalized customer service experience to users. Intelligent customer service with artificial intelligence technology as its core plays an important role in improving user experience.
The essence of intelligent customer service is to fully understand the user's intentions and accurately find the matching knowledge points in the knowledge base to answer user questions or provide solutions. The entire intelligent customer service is built around the user's problem. The problem similarity calculation is the core technology of almost all links, such as intelligent customer service knowledge base construction, online problem matching, full-link data operation, etc. The construction of its related data will inevitably promote the progress of the entire customer service industry.
“In the financial sector, the challenge itself is very broad and comes from many sources,†said Yan Yuan, vice president and chief data scientist of Ant Financial, “To solve these problems, we have developed a series of artificial intelligence algorithms.â€
At the core of the development algorithm, that is, in the process of building an artificial intelligence engine, Ant Financial also encountered many challenges, such as enabling the machine learning engine to quickly re-judge, how to face very small samples to achieve learning and prediction.
"The first job we did here is to be on the smart customer service. I hope that through the ability of multiple rounds of dialogue, Alipay will be smarter and understand the meaning behind the dialogue." Painter added.
Zhang Jiaxing, senior algorithm expert of Ant Financial, said in the elaboration of the game title, "This game is very simple to define. It is to give you two sentences. Let you judge whether the semantics of these two sentences are consistent. For example, a sentence It is 'How to repay the flower buds', and the other sentence is 'How to pay back the flowers.' These two sentences have many different words, but they have the same meaning."
The significance of this in intelligent customer service is very important, but the back is not simple.
First of all, the most important point of customer service is to understand the user's intentions, and then match the relevant knowledge points in the knowledge base. At present, the most mainstream matching practice in the industry is to calculate the similarity of the user's problem and all the problems in the knowledge base one by one, and regard the most similar problem as the user's intention, and then return its answer to the user.
In addition, the similarity calculation is also meaningful for all aspects of the entire customer service. For example, the knowledge base is actually based on automatic text-based mining. The automatic mining algorithm of this text is also based on calculating the similarity between any two texts or the distance between two high-dimensional spaces.
In all aspects of data processing, including search recommendations, dialogue and other areas, similarity calculations are also very core and fundamental issues.
At present, the biggest challenge and problem facing NLP similarity calculation lies in the fact that there are a large number of very diverse problems in intelligent customer service, including different dialects, spoken language and various expressions, and the matching of problems. Quite high demand. However, the progress of NLP is mainly concentrated in sentiment analysis and translation. In these scenarios, most of the time it is in a relatively fixed data set, which can get good results on a relatively simple and clean data set.
"The challenge here is actually very simple. It is because people have too many possibilities when they express the same meaning. There are various different opinions." Zhang Jiaxing also gave his own solution ideas:
The first is to pay attention to the expression of the problem and the expression of different levels. The natural language that people see is also a form of expression, and many things can be done on this basis. For example, based on the deep learning model, the sentence is turned into a vector, and the distance between the vectors is calculated. It is also possible to establish a grammar tree by dependency analysis, and establish a model matching at the tree level; or the sentence can be truly understood to the semantics. Hierarchy, in this way, matching the semantics of any two sentences may achieve better results. This is an attempt in practice. I hope everyone can try more ways.
The second point is that machine learning relies on big data. This question is for the customer service field. I hope that you can explore more data in the field and use some machine learning methods to combine the things learned in other fields with the data in the customer service field. Come up and finally achieve better results.
In fact, whether the sample information is asymmetry caused by the change of black production or the weak adaptability of the model caused by the diversity of customer service data, these problems must solve the "small data" problem. In theory, we often say how to do with big data, but it is difficult to achieve such an ideal environment in the actual engineering process. Whether it is migration learning or unsupervised learning, how to achieve higher precision with less samples, faster recognition rate, and improved generalization of machine learning are all challenging tasks in the next stage.
This series of propositions partly constitutes the most difficult problem to be solved in the new financial field.
Gasoline Gensets,Silent Gasoline Gensets,Homemade Electrical Generator,High Pitch Sound Generator
Wuxi Doton Power , http://www.dotonpower.com