Why is knowledge-driven artificial intelligence gone?

But the ideal is full and the reality is skinny. In a sense, artificial intelligence at this stage belongs to weak artificial intelligence, that is, the stage of "mental retardation".

At present, artificial intelligence is based entirely on data reasoning. First, I can't understand human emotions, and I can't communicate with people normally or even deeply; Second, it can't cope with the decision-making and planning problems in complex scenes.

At present, the landing of artificial intelligence is the intelligence of fast feedback and single scene perception. Just like a newborn child, you can see and hear, but it takes a long time to understand and comprehend.

So the next important task of artificial intelligence and data intelligence is how to make machines or software have brains.

There are four steps in the development of data intelligence or artificial intelligence: expert intelligence, data-driven AI, big data-driven AI and big knowledge-driven AI.

The first step is expert information. In the 1960s, many experts regularized logic with some rules or symbolic logic, hoping to replace work with automated processes as much as possible.

Indeed, this is valuable in many scenarios and can improve labor efficiency, but there are so many experts who use rules to solve this problem. The written rules are very complicated and their applicability is very limited.

Therefore, the knowledge base of the expert database has been stagnant for a long time.

Later, machine learning algorithm was introduced, and we made great breakthroughs in many decimal calculations, which helped us to make predictions in many scenarios. However, this kind of threshold is relatively high and requires characteristic engineering. In addition, different algorithms need to be selected, which is a relatively small application scenario.

Until the emergence of 20 1 1 deep learning, it was a direct end-to-end product without knowing the algorithm. The specific parameters inside do not need to be designed, and the result goes directly to artificial intelligence driven by big data.

Recently, it has been mentioned that "making great efforts to create miracles" is not the progress of algorithms, but the progress of computing power in data use and artificial intelligence driven by big data.

But this kind of deep learning has a big bottleneck.

First, the feature problem can't explain what works and can't be applied in many key scenarios.

Second, the generalization problem, the so-called big data-driven artificial intelligence, needs to learn a lot of data in order to learn a good model. This is different from people. As a creature, man has a set of theoretical framework and a set of common sense maps. Under the framework of common sense diagram, a few diagrams can produce very good generalization effect.

Artificial intelligence driven by big knowledge is coming soon.

Therefore, we believe that the next generation of artificial intelligence will definitely appear in the next five to ten years, and many people mention artificial intelligence driven by big knowledge.

The big knowledge-driven artificial intelligence, the challenge we face, is first of all big.

The traditional knowledge base is limited, but now there is a lot of knowledge. The cloud on the database, including human body data and Internet of Things data, can build a knowledge map.

Because most of the core personnel of Qunzhi come from Microsoft Research Institute, they used to do search and data mining in Microsoft Research Institute.

Search is the first large-scale commercial application of artificial intelligence and data. Many people will ask, which market is larger than NLP and knowledge map?

In fact, it is obvious that the knowledge map has been proved because of natural language. Search is a huge market, and the core technology of search itself is to build a super-large-scale knowledge map with open data, understand documents and users' queries, and finally achieve accurate matching between queries and documents.

From the search point of view, it is a general query decision engine.

After we came out from Microsoft, we were wondering whether we could apply this technology to enterprises and industries, and how to combine general knowledge map with industry knowledge map to solve industry problems. This is our initial intention to take a look at the wisdom of the group.

Solve the core problems of artificial intelligence application and create a general knowledge map.

The core problem is how to extract knowledge from massive multi-source heterogeneous data, establish relationships, disambiguate and fuse different data and construct knowledge. This is the first problem of knowledge construction.

The second question is how to understand semantics. For example, at present, most human-computer dialogues, siri, and human-computer dialogues in various fields do not understand semantics, which is the difficulty at this stage. I firmly believe that the rapid development of natural language processing technology, especially semantic understanding technology, will form a huge breakthrough. Just like the breakthrough of visual technology in previous years.

The third is the issue of knowledge empowerment. With great knowledge, how can it be combined with biotechnology and language technology? We have a general knowledge map, how to train and recognize a specific engine, and the goal is to make a recognition engine based on four to five pictures. The situation now is that there are still thousands of pictures to be marked.

Our industry is similar to the search industry. What we need to do at this stage is to land on the intelligent platform.

The intelligent decision-making platform includes four basic products, which respectively solve the problems of data fusion and fast data closed loop, knowledge map construction, correlation analysis and AI model training, that is, decision-making problems. This is a set of products or methods based on our team's years of experience.

Specifically, the auxiliary decision system of intelligent decision-making platform consists of four parts: perception, understanding, analysis and decision.

First of all, our products solve the problem of extracting structured information from multi-source heterogeneous data including Internet of Things data and public industry data to form a general knowledge map. In this process, we need to use natural language understanding and knowledge extraction to build a knowledge map.

When you have a knowledge map and use basic machine learning products, you can lower the threshold for many industry personnel and practitioners to use artificial intelligence products, and you can quickly build decision-making products.

With decision-making products, products do not replace people, but help people understand data in the decision-making process and quickly use people's decision-making ability in human-computer interaction.

After constructing such a framework system, the common sense map combines the data of specific segments such as public security and media to empower the industry.

Next, let's talk about the application scenarios we have done in several segments.

Application of knowledge map in various industries

The application of the first intelligent decision-making platform in public security industry.

Because the public security has data, it can be commercialized. The current function is biased towards post-analysis. After data classification, if there are cases, just use atlas, data in the middle of the station or business in the middle of the station to catch criminals.

Another scene is a bank.

Artificial intelligence really liberates people from simple work. A single intelligent expert examination system was successfully launched in a large domestic bank last year, which combined artificial intelligence technology with international document business for the first time and achieved a breakthrough in the field of artificial intelligence in document business from scratch.

The effect is very obvious. In terms of quantity, the original undergraduate and graduate students need 2000 senior talents, but now they only need 100 ordinary talents. Moreover, the efficiency of examining documents is greatly improved, and the lag phenomenon is basically eliminated. For a large number of bills, the key elements are processed and matched with the knowledge map. This technology is in great demand in insurance, banking and securities industries.

The other is anti-money laundering products.

As we all know, anti-money laundering is the blood of the whole finance, and all financial crimes are related to funds, such as corruption, bribery and terrorist financing.

Therefore, the whole financial system, like the requirements of national financial supervision, must be based on the anti-money laundering system.

The purpose is clear.

First, prevention is achieved by manually adding rules. For example, this person didn't lend money before, and suddenly he made a huge sum of money, which may be a problem. For example, if you are in the Apple business, you suddenly make a huge sum of money for people in the mining business and need to report it to the People's Bank of China.

These possible problems must be checked manually first. For example, a bank has nearly 80,000 possible suspects every day, but it needs to be manually turned into 200.

The problem with the current mode is that the response is very slow. If we use our model, first of all, we can quickly identify whether it is a money laundering transaction. Second, real-time delivery can be realized, which can effectively put an end to financial crimes.

The above is our understanding and some landing scenes.

Every technological revolution will bring anxiety, but every time the result is that people live happier. Next, the new chapter of artificial intelligence must be people-oriented. The most important problem is to make machines look like people, better understand people, communicate with people, make people become superhuman and enlarge their skills.

Thank you all.

This website only provides storage services, and all contents are published by users. If you find harmful or infringing content, please click Report.

Expand the rest of the content