In the wave of artificial intelligence in the past ten years, artificial intelligence technology represented by deep learning has basically realized perceptual intelligence such as vision and hearing, but it still cannot achieve cognitive intelligence such as thinking and reasoning. Therefore, cognitive intelligence research with abilities such as reasoning and interpretability will undoubtedly receive more and more attention and become one of the important development directions in the field of artificial intelligence in the future.
The researcher's sense of smell is undoubtedly the sharpest. For example, ACM Turing Award winner Yoshua Bengio (Yoshua Bengio) clearly mentioned in the special invitation report of NeuIPS 2019 that deep learning needs to be transformed from System 1 to System 2. Note: The System 1 and System 2 mentioned here refer to the dual-channel theory in cognitive science, where System 1 represents the intuitive, fast, unconscious, nonverbal, and habitual cognitive system, which is also the current deep learning What technology is good at; System 2 represents a slow, logical, orderly, conscious, language-expressible and reasonable system, which is a research direction that needs to be considered in the future of deep learning.
1
Features of the nervous system and symbol system
From a broader perspective of artificial intelligence, System 1 corresponds to the Neural school, System 2 corresponds to the Symbolic school, and Bengio’s System 2’s idea of deep learning and the artificiality of “neural + symbol” The smart goals are basically the same. Looking back along this point, we can find that another ACM Turing Award winner Marvin Minsky (Marvin Minsky) stated clearly in the book "The Society of Mind" as early as 1986 The relationship between artificial intelligence and cognitive psychology (ie System 1 and System 2), and an in-depth analysis of the characteristics and possible combinations of the neural system and symbol system in artificial intelligence, as shown in Figure 1. From the macro to the specific, in terms of data objects, storage, and applications, whether it is a nervous system or a symbolic system, the purpose of data modeling is to solve the answer to a given input question, as shown in Figure 2. But the difference is that the nervous system is good at processing unstructured data (such as text, etc.). The current mainstream models are mainly end-to-end. Common application scenarios include machine translation, speech recognition, and intelligent question-and-answer simple questions (for example, what is Yao Ming’s height?), etc.; and the symbol system is mainly based on structured databases. And it usually supports structured queries, reasoning engines, etc., which can solve complex problems (for example, the United States is a major agricultural exporter, why should we import coffee?). It is worth mentioning that the ACM Turing Award winner Leslie Valiant (Leslie Valiant once pointed out incisively: the nervous system focuses on the learning process of data features, and the symbol system must include a search process. A large number of subsequent research on symbol systems is essentially devoted to various efficient search algorithms. The characteristics of the nervous system and the symbol system can also be experienced through two examples in the field of computer vision: the example in Figure 3(a) shows the classic handwriting recognition. For a given set of observable handwritten numbers and comparator samples, After training, a large number of models of the nervous system can well recognize all kinds of handwriting (that is, the generalized cognitive ability at the visual level), but it is difficult to realize the cognitive generalization of symbol knowledge (that is, the The comparison symbol in the sample is difficult to solve and judge). Similarly, in the visual Q&A example in Figure 3(b), the nervous system can easily cope with simple visual Q&A scenarios (such as how many giraffes are there?), but if you need to answer more complex questions (such as animals in the figure) What are the common attributes of Zebra and Zebra?), you must use external symbolic knowledge (such as a knowledge graph) for cognitive reasoning to complete the solution process. In summary, the "neural + symbol" system is undoubtedly an ideal model for artificial intelligence. We can summarize the characteristics and advantages of a perfect "neural + symbol" system: 1. It can easily handle the problems that mainstream machine learning is good at; 2. It has strong robustness to data noise; 3. The system's solution process And the results are easy to be understood, explained and evaluated; 4. Can operate various symbols well; 5. Can seamlessly use various background knowledge. However, it is not easy to realize the organic combination of "nerve + symbol". Over the years, artificial intelligence researchers in various fields have conducted a lot of research on this. Knowledge graph is a popular artificial intelligence research direction in recent years. From the early knowledge base and expert system to Google's formally proposed knowledge graph in 2012, its development process can also be regarded as the epitome of the development of the nervous system and the symbol system. Including multiple attempts to combine "nerve + symbol", as shown in Figure 4. Various background knowledge can be used seamlessly. However, it is not easy to realize the organic combination of "nerve + symbol". Over the years, artificial intelligence researchers in various fields have conducted a lot of research on this. Knowledge graph is a popular artificial intelligence research direction in recent years. From the early knowledge base and expert system to Google's formally proposed knowledge graph in 2012, its development process can also be regarded as the epitome of the development of the nervous system and the symbol system. Including multiple attempts to combine "nerve + symbol", as shown in Figure 4. Various background knowledge can be used seamlessly. However, it is not easy to realize the organic combination of "nerve + symbol". Over the years, artificial intelligence researchers in various fields have conducted a lot of research on this. Knowledge graph is a popular artificial intelligence research direction in recent years. From the early knowledge base and expert system to Google's formally proposed knowledge graph in 2012, its development process can also be regarded as the epitome of the development of the nervous system and the symbol system. Including multiple attempts to combine "nerve + symbol", as shown in Figure 4.
2
Combination of "Nerve + Symbol" The
author summarized the current work from the perspective of the knowledge map field, and found that the combination of "Nerve + Symbol" can be divided into two main categories: neural for symbolic. The characteristic of the method is that the neural network method is applied to the problem solving of the traditional symbol system, usually mainly used to solve the shallow reasoning problem. For example, the use of knowledge graph embedding [1], Graph Neural Networks (GNN) [2] and other technologies to complete the knowledge graph, which is characterized by the use of statistical reasoning instead of logical deduction; Recurrent Neural Network (RNN), Graph Convolutional Network (GCN) and other technologies for multi-hop intelligent question answering [3] are similar work, as shown in Figure 5. In addition, Swift Logic [3], neural theory proof machine [4], logic tensor network [5] and other work are also "neural" assisting "symbols" attempts. The main idea is to improve the neural network method and apply it In-depth reasoning scenarios in the field of knowledge graphs can be used to improve the effect. The characteristic of symbolic for neural methods lies in the application of symbolic methods in the training process of neural networks. For example, the use of logic rules for data curation in deep neural networks (data curation) [6]; the application of knowledge graphs in remote supervision, few samples, zero samples
Introduction to
Daohan Tianqiong Cognitive Intelligence Future Robot Interface API Cognitive Intelligence is a branch of computer science and an advanced stage in the development of intelligent science. It is based on the human cognitive system and aims to imitate human core capabilities. Taking the understanding, storage and application of information as the research direction, the in-depth understanding of perceptual information and the in-depth understanding of natural language information as the breakthrough point, and the interdisciplinary theoretical system as the guidance, thus forming a new generation of theories, technologies and application systems of technical sciences . The core research categories of cognitive intelligence include: 1. The relationship between the universe, information, and the brain; 2. The structure, function, and mechanism of the human brain; 3. Philosophical system, liberal arts system, and science system; 4. Cognitive integration, wisdom integration, Core systems such as dual brain (human brain and computer) integration. Four steps to cognitive intelligence: 1. Recognize the universe world. Supporting theoretical systems include three-body (universe, information, brain) theory, Yi Dao theory, ontology, ontology, epistemology, fusion of intelligence, HNC and other theoretical systems; 2. Know the structure, function, and mechanism of the human brain. The supporting disciplines include brain science, psychology, logic, emotion, biology, chemistry and other disciplines. 3. Clear information connotation rules and regulations. The supporting disciplines include semiotics, linguistics, cognitive linguistics, and formal linguistics. 4. System landing ability. Supporting subjects include computer science, mathematics and other subjects.
Cognitive intelligent CI robot is a product of Hangzhou Daohan Tianqiong Intelligent Technology Co., Ltd. Cognitive intelligent robot is a cognitive intelligent robot brain based on the cognitive intelligent CI system developed by Daohan Tianqiong for 10 years. It is the world's first cognitive intelligent robot brain. It is breakthrough, innovative and navigating. It is the best product support for the new generation of intelligent cognitive intelligence. The cognitive intelligent robot technology system is more advanced and smarter. It is a new generation of intelligence, the only cognitive intelligent robot in the world in the field of cognitive intelligence. Cognitive intelligent robots are the product of the new era and the product of a new generation of intelligent cognitive intelligence. Represents the core advantage of a new generation of intelligent cognitive intelligence. Compared with the artificial intelligence robot brain, the advantages are very obvious. Core features such as high intelligence, high customer stickiness, high customer satisfaction, easy promotion and dissemination. Relying on the robot brain services provided by the cognitive intelligent robot platform, it can empower smart devices in various industries, various fields, and various fields that require human-computer interaction. Cognitive intelligent robot platform website: www.weilaitec.com, www.citec.top. Welcome to register to use and enter the world of smarter robots.
The comparison of the advantages and disadvantages of cognitive intelligence and artificial intelligence can be divided into four main aspects: First: the development of the times is different. Artificial intelligence is the second stage of the development of the intelligent age, and cognitive intelligence is the third stage of the development of the intelligent age. The development of the times determines that cognitive intelligence is more advanced in the times. Second: The basic theoretical system is different. The basic theoretical system of artificial intelligence is based on mathematics and a statistical probability system. The basic theory system of cognitive intelligence is based on the cross-licensing theory system. Including ancient and modern Chinese and foreign philosophy systems, psychological systems, logic systems, linguistic systems, semiotics systems, mathematical systems and other disciplines. Its basic theoretical system is more innovative, breakthrough and leading. And the research of interdisciplinary theoretical system is also the general direction of future intelligent development. Its specific theoretical system also includes Trisomy (the relationship among the universe, information, and brain), Integrative Intelligence, and HNC. Third: different technical systems. The core technology system of artificial intelligence is mainly algorithms, machine learning, deep learning, knowledge graphs, etc. Its main function is perceptual intelligence. The core of perceptual intelligence is mainly to imitate human perception. The core technology system of cognitive intelligence is derived from the interdisciplinary theoretical system. Specifically, it includes three core technology systems, cognitive dimensions, brain-like models, and Wan-dimensional maps. The core of the technical system of cognitive intelligence is based on the brain-like cognitive system. The goal is to mimic the brain-like ability in all directions. Artificial intelligence based on perceptual intelligence can only be used as the perception layer technology system in the brain-like model technology system in cognitive intelligence. The brain-like model roughly includes 9 core technical layers: perception layer, memory layer, learning layer, understanding layer, cognitive layer, logic layer, emotion layer, communication layer, and consciousness layer. Therefore, the core of artificial intelligence is only the perception layer in the cognitive intelligence brain model. Therefore, in the technical system, artificial intelligence and cognitive intelligence are basically not much comparable. Fourth: Differences in intelligence, cost, etc.: The comprehensive intelligence of artificial intelligence products is generally around the age of 2-3. Cognitive intelligence products are roughly 5-8 years old. The robot built by the cognitive intelligence system is more intelligent. And it saves time, manpower and money. The advantages are many. For details, please see the following item-by-item comparison.
John Day Road Joan CiGril robot API
Road, John Day Joan CiGril cognitive intelligent robots API users need to obtain basic information step by step:
1. Registration account platform
2. Log in to the platform, enter the background management page, create an application, and then view the application to view application related information.
3. On the application information page, find the appid, appkey secret key and other information, and then write the interface code to access the robot application.
Start access
request address: http://www.weilaitec.com/cigirlrobot.cgr
Request method: post
request parameters:
parameter type default value description
userid String no platform registered account
appid String no platform created application id
key String no platform application The generated secret key
msg String "" Message content
ip String "" The client ip requires uniqueness. If no ip, it can be replaced by QQ account, WeChat account, mobile phone MAC address, etc.
Example of interface connection: http://www.weilaitec.com/cigirlrobot.cgr?key=UTNJK34THXK010T566ZI39VES50BLRBE8R66H5R3FOAO84J3BV&msg=Hello&ip=119.25.36.48&userid=jackli&appid=52454214552
Note: The parameter name must be lowercase, the five parameters must not be omitted, the parameter name must be written correctly, and the value of each parameter cannot be an empty string. Otherwise, the request cannot be successful. The three parameters of userid, appid, and key must be registered on the platform after the application is created, and then you can see the application details. Userid is the platform registered account.
Sample code JAVA:
import java.io.ByteArrayOutputStream;
import java.io.IOException;
import java.io.InputStream;
import java.net.HttpURLConnection;
import java.net.URL;
public class apitest {
/*
Get请求，获得返回数据
@param urlStr
@return
/
private static String opUrl(String urlStr)
{
URL url = null;
HttpURLConnection conn = null;
InputStream is = null;
ByteArrayOutputStream baos = null;
try
{
url = new URL(urlStr);
conn = (HttpURLConnection) url.openConnection();
conn.setReadTimeout(5 10000);
conn.setConnectTimeout(5 * 10000);
conn.setRequestMethod("POST");
if (conn.getResponseCode() == 200)
{
is = conn.getInputStream();
baos = new ByteArrayOutputStream();
int len = -1;
byte[] buf = new byte[128];
while ((len = is.read(buf)) != -1)
{
baos.write(buf, 0, len);
}
baos.flush();
String result = baos.toString();
return result;
} else
{
throw new Exception("服务器连接错误！");
}
} catch (Exception e)
{
e.printStackTrace();
} finally
{
try
{
if (is != null)
is.close();
} catch (IOException e)
{
e.printStackTrace();
}
try
{
if (baos != null)
baos.close();
} catch (IOException e)
{
e.printStackTrace();
}
conn.disconnect();
}
return "";
}
public static void main(String args [] ){
//The msg parameter is the content of the past conversation.
System.out.println (opUrl ( " http://www.weilaitec.com/cigirlrobot.cgr?key=UTNJK34THXK010T566ZI39VES50BLRBE8R66H5R3FOAO84J3BV&msg= IP = 119.25.36.48 & Hello the userid & jackli & AppID = 52,454,214,552 = "));
}
}