TW201802705A - System having artificial intelligence - Google Patents

Abstract

The present invention provides a system (1) having a first artificial intelligence module (30) that includes a function to assist in or autonomously perform information exchange in an open environment, and a second artificial intelligence module (50) for monitoring the operation of the first artificial intelligence module (30) and autonomously performing a process that accompanies information exchange in a closed environment. A system (1) can be provided that enables use outside (external) resources via the Internet or the like in order for the first artificial intelligence module (30) to learn or make inferences, and at the same time, to protect against illegal accesses from the outside.

Description

System with artificial intelligence

The invention relates to a system containing artificial intelligence.

Japanese Patent Laid-Open Publication No. 2014-222517 describes that a device and a user can easily communicate with each other and use local services and / or remote services more efficiently. The intelligent automatic assistance system of this document is to call an external service when using natural language dialogue and when it is suitable for obtaining information or performing various actions related to the user based on the unified dialogue. It is realized using any of a variety of platforms such as websites, emails, and smartphones, or all combinations thereof. Based on a collection of fields and tasks that are given a relationship to each other, additional functionality is provided to supply power through external services that the system can talk to.

Japanese Patent Laid-Open Publication No. 2008-165824 provides a method and system for assessing the risk of improper transaction of electronic commerce between customers and sellers. In this document, an order for e-commerce or e-shopping is received, and the risk score of return is determined along with the risk level of each order. Data verification, highly predictive artificial intelligence model matching, network data synthesis, and negative file check are the most important factors used to test the risk of improper trading. Other analyses also include this transaction's mishandling of known past Easy comparisons and unusual speed patterns, name and domicile changes, and retrieval of a transaction history database to identify known fraudsters. Micro-callable to adapt the services provided by the system to new or changing improper transaction models.

[Prior technical literature] [Patent Literature]

[Patent Document 1] Japanese Patent Publication No. 2008-165824

Artificial intelligence (AI) is used for various processing or functions such as execution, assisting user interface, and discovery of inappropriate behavior. It is expected that artificial intelligence (artificial intelligence technology) will be further applied to various types of Kind of field. Artificial intelligence, for example, presents a computer system with functions of human intelligence such as learning / inference / judgment. A system that is useful as a human supplementary has been developed, for example, a system that solves difficult problems in specialized fields such as medicine / science / engineering at the same level as human experts in its field, and inter-lingual The system of translation, the system of operating robots that perform intelligent operations, and the system of discovering useful information from websites are not limited to these.

It is important for artificial intelligence to operate normally without improper access in the case of replacing human beings with intelligent actions such as language understanding or inference, and problem solving. In order for artificial intelligence to learn and infer, it is necessary to use external (external) resources via the Internet, etc. On the contrary, it is increasingly difficult to defend against improper access from the outside.

One aspect of the present invention is a system having: a first artificial intelligence module including a function to assist or autonomously exchange information in an open environment; and a second artificial intelligence module to monitor the first artificial intelligence The operation of the module performs processing accompanied by information exchange in a closed environment. The first artificial intelligence module (front side artificial intelligence) is based on information containing external resources in an open environment, and can assist or autonomously perform the information exchange desired by users or hosts. The second artificial intelligence module (back side artificial intelligence) operates in a closed environment and monitors the operation of the first artificial intelligence module without being exposed to improper access from the outside. In this system, the second artificial intelligence module can indirectly grasp the information exchange desired by the user or the host, and autonomously perform the information exchange-based processing in a closed and secure environment. Therefore, the system containing the first artificial intelligence module and the second artificial intelligence module is in a safe state in which the leakage of information or misoperation due to improper access, etc., and the obstacles that have been worried about have been prevented , It can respond to user or host requests while effectively utilizing many resources provided in an open environment.

The open environment may also include a user interface, a first network connection interface for connecting to an open network service, such as the Internet, and an exchange with the user interface or the first network connection interface. Of the first memory. The closed environment may also include: the second memory that the first artificial intelligence module cannot access, and the second artificial intelligence module may also include: the information that will be requested along with the information exchange and will be stored in the second The secret information of the memory is encrypted for the function of the first memory. Use card information, account information, and other personal information about payment, user ID, password, and other personal information as secret information to use the second person Intelligent modules can be safely managed.

The closed environment may include a second network connection interface for connecting to a closed network service separated from an open network such as a home LAN, a car LAN, a factory LAN, and a store LAN. As a home gateway, a car gateway and a factory gateway, the first artificial intelligence module is used for access from the open network, and the second artificial intelligence module is used for the separated network. Group to correspond. With this system, the connection to the external network can be answered while guaranteeing the security of the separated network. The second artificial intelligence module may also include a function of monitoring the abnormality of the first artificial intelligence module.

This system may also have a first OS that supports an open environment and is the first OS operating by the first artificial intelligence module; and a second OS that supports the closed environment and is the second OS that operates by the second artificial intelligence module. The second OS operating in a protected environment separated from the first OS can ensure the safety of the second artificial intelligence module.

A typical architecture having a memory including a secured area and a non-secured area, a processor unit including a support mechanism that prohibits access to the secured area, and a hypervisor operating on the processor unit may be used. The hypervisor includes: a mechanism that permits access to the second security zone and the non-security zone and prohibits access to the open environment; and a permission that the first OS accesses the non-security zone and the non-security device and prohibits use of the support mechanism North Korean security zone access agency.

The hypervisor may further include: virtualizing hardware (device) included in the processor unit and providing the virtualization support mechanism to the first OS and the second OS. The processor unit may also include: 1 or a plurality of main processor units (CPUs); and 1 or a plurality of graphics processing processor units (GPUs). The hypervisor may also include: virtualization support functions for at least one CPU and at least one GPU.

The first artificial intelligence module may also include: a plurality of mechanical learning modules; a plurality of mechanical learning materials used in each mechanical learning module; a communication unit between mechanical learning modules that performs communication between the plurality of mechanical learning modules ; And a separate unit that suppresses the mixing of multiple machine learning materials. As a mechanical learning module (technique), proposals include ID3 (Iterative Dichotomiser 3) or random forest (Random Forest, Randomized trees) and other decision tree learning, related rule learning, and various neural networks (ANN (Artificial Neural Network), DNN (Deep Neural Network), CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), etc.) , Genetic Programming (GP), Inductive Logic Programming (ILP), Support Vector Machine (SVM), Clustering, Bayesian Network, etc. Each has its own advantages.

As an artificial intelligence module of this system, a mechanical learning integration platform including a plurality of mechanical learning modules, a plurality of mechanical learning materials, a communication unit between modules, and a separate unit can also be used. Become an artificial intelligence module that can provide a suitable combination of several types of mechanical learning modules, and provide an artificial intelligence module suitable for personalization. In particular, the artificial intelligence module including the mechanical learning integration platform is a first artificial intelligence module that is suitable for use in a variety of purposes. The second artificial intelligence module may also include a machine learning integration platform.

The first artificial intelligence module may also include: a unit that judges the failure and success of the response to the user; and automatically collects data from an open environment to automatically construct a personal database of data that will successfully guide the response to the user single yuan. Personalization of artificial intelligence can be performed autonomously. The artificial intelligence module with a unified platform can also include: automatically collecting mechanical learning modules and mechanical learning data from the open environment to successfully guide the user's response and installing them in the installation of the first artificial intelligence module unit.

The first artificial intelligence module and the second artificial intelligence module can also be constructed on the server, but are divided into an open environment and a closed environment. In order to ensure the operation of the first artificial intelligence module according to the second artificial intelligence module For traceability, the control module of the first artificial intelligence module and the second artificial intelligence module should be equipped on a single chip or chipset.

An example of this system includes a personal assistance system connected to the user interface of the first artificial intelligence module, such as a PDP, a PDA, a smart phone, and the like. Two other examples of the system are a home gateway system including a user interface connected to the first artificial intelligence module and a home LAN and / or device control interface connected to the second artificial intelligence module. One of the other examples of this system is an on-vehicle gateway system including a user interface connected to the first artificial intelligence module and a vehicle-mounted LAN and / or device control interface connected to the second artificial intelligence module. One of other examples of this system is an IoT gateway that includes an Internet interface connected to the first artificial intelligence module and a factory LAN and / or device interface connected to the second artificial intelligence module.

1‧‧‧ system

2‧‧‧ users

3.71‧‧‧Smartphone

5‧‧‧ Cloud Service

6‧‧‧ Security Mode

7‧‧‧ non-security mode

8‧‧‧ closed environment

9‧‧‧ Non-secure environment

10‧‧‧hardware platform

11‧‧‧ processor unit

12‧‧‧ Support Agency

12a, 12b‧‧‧Memory space

13, 13a, 13b‧‧‧Memory

14‧‧‧user interface

15, 17‧‧‧ web interface

16, 18‧‧‧ connection interface

20‧‧‧ Hyper Manager

20v‧‧‧Virtualization Support Agency

21‧‧‧ Non-Security OS

21a ~ 21d‧‧‧Function OS

22‧‧‧Security OS

30‧‧‧ conventional AI

38‧‧‧ Response Failure and Success Unit

39‧‧‧ Unit of Personal Database

41 ~ 44, 51 ~ 55‧‧‧Function modules

41a‧‧‧Sound Synthesis / Conversation Agent

41b‧‧‧Voice recognition function

41c‧‧‧Voice dialogue platform

43a‧‧‧ Store App

43b‧‧‧Internet Services

43c‧‧‧Programming

47‧‧‧ Non-Security App

49b‧‧‧navigation system

49c‧‧‧security system

49d‧‧‧Advanced Operation Support System

50‧‧‧Security AI

57‧‧‧Security App

58‧‧‧Function of abnormal movement

59‧‧‧Monitoring function

73‧‧‧Home Gateway System

75‧‧‧Car gateway

80‧‧‧AI integration platform

81a ~ 81d‧‧‧Learning Module

81v, 81x, 81y‧‧‧ engines

82a ~ 82d‧‧‧Learning materials

83‧‧‧ separation unit

84‧‧‧Sound processing function

85‧‧‧ Inter-module communication unit

87a‧‧‧data exploration

87b‧‧‧ Data Cluster

87c‧‧‧Web Search

87d‧‧‧Noise Filter

87e‧‧‧Mission Agent

87f‧‧‧CV

87g‧‧‧Type / vocabulary analysis

87h‧‧‧Sensor I / F

88‧‧‧judgment unit

89‧‧‧ Self-discipline Expansion Unit

91‧‧‧Early Learning

92‧‧‧ Dynamic Learning

101‧‧‧ main processor unit

102‧‧‧Graphics Processor Unit

105‧‧‧Computer Resources

Figure 1 contains a block diagram of a conventional AI and security AI system.

FIG. 2 is a block diagram of a CPU and a system including a GPU.

Figure 3 shows a diagram of an example applicable to a personal assistance system.

FIG. 4 shows a diagram corresponding to an inappropriate state.

FIG. 5 is a diagram showing an example applicable to a home gateway system.

Figure 6 shows a block diagram of an AI integration platform.

Figure 7 diagram of deep learning.

FIG. 8 shows a graph of initial learning and dynamic learning.

FIG. 9 is a diagram showing an example of initial learning and dynamic learning.

FIG. 10 is a diagram showing another example of initial learning and dynamic learning.

FIG. 11 is a diagram showing another example of initial learning and dynamic learning.

FIG. 12 shows an example of dynamic learning using dialogue.

FIG. 13 shows a dynamic learning example of voice recognition.

Figure 1 shows an example of a system with two artificial intelligences. This system 1 is built into personal assistant terminals, such as computer systems for smart phones, notebook computers, home appliances, industrial machinery, and autos. It is typically loaded into a SoC (system-on-chip). system. The system 1 includes: a hardware platform (HardWare flatform; H / W flatform) 10 including a processor unit 11; a hypervisor 20 operating thereon; a first OS operating in a virtualized environment provided by the hypervisor 20 ( Non-secure OS, non-secure OS, general purpose OS, multi-function OS) 21; and 2nd OS (secure OS) 22.

Although the processor unit 11 provides a plurality of execution modes (execution environments), in this embodiment, it is supposed to be realized by a support function (memory protection function) that sets a security area in the memory and prohibits access to the security area. Security mode 6 and low security mode (hereinafter referred to as non-security) (Full mode, non-secure mode, or normal mode) 7 and 2 execution modes. In the non-security mode 7, access to the security area of the memory by the support function of the processor unit 11 is prohibited.

System 1 has: a trusted application group (security app) 57 that is safe on security OS 22 but operates in a secure environment (Secure World, Closed World) 8 operating in security mode 6; and in non-security The multi-function OS operating in mode 7 is a client application group (non-security app) 47 which is open on non-security OS 21 but operates for non-security environment (Non-Secure World, Normal World, OpenWorld) 9.

The non-security app 47 series includes: the first artificial intelligence module (conventional AI, pre-AI, open AI) 30 operating on the multifunctional OS21; and various functional modules (application) 41 supported by the conventional AI30. 44. The security application 57 includes: the second artificial intelligence module (security AI, rear AI, closed AI) 50 that operates in the second OS22 in a closed environment; and various functional modules (applications) 51 supported by the security AI50. 55. System 1 may further be provided with a functional module that operates on the multifunctional OS21 independently from the conventional AI30, and may further be provided with a functional module that operates on the secure OS22 independently from the security AI50.

The functional modules supported by the conventional AI 30 include a sound processing function 41, an image processing function 42, various service functions 43, and various search functions 44. The voice processing function 41 includes, for example, voice recognition, voice synthesis, and dialogue functions. The image processing function 42 includes an authentication function through an image, such as a face authentication, a recognition function of the surrounding environment, an indoor, outdoor, place, day and night recognition function. The service function 43 includes an Internet connection service in an open environment such as a shop app or the Internet. Retrieval function 44 includes retrieving Information or the function of retrieving external information via the network.

The function modules and functions supported by the conventional AI30 are not limited above. Furthermore, the number and type of function modules can be increased or decreased by personalization and / or customization. To this end, the conventional AI 30 includes: a unit 38 for judging the failure and success of the response to the user; and automatically collecting data (information) from an open environment to automatically guide the user's response to construct a personal database Unit 39.

The functional modules supported by the security AI50 include: the function of making payment (checkout) (payment function unit) 51; personal information such as card information (personal information) and management of other secret information or secret information (the following includes Function (unit) 52 for secret information); for monitoring in homes, vehicles, factories, etc., in particular, to control sensors for obtaining better information for personal or other security (confidential) management to Function (unit) 53 for obtaining data; control of switches and the like of various devices installed in homes, vehicles, factories, etc., especially control of devices for preventing crimes, devices that affect security or safety A function (unit) 54; and a function (unit) 55 that encrypts the secret information.

An example of the hypervisor 20 is a hypervisor corresponding to a support organization that controls writing to a secure area of a memory and is a FOXvisor provided by the applicant of the present application. The other example is a hypervisor corresponding to a support organization and having a communication function between mechanical learning described later, and a VAIOX provided by the applicant. These are bare metal type (type I) hypervisors.

As shown in FIG. 1, the VAIOX 20 series can use a processor (processor unit) 11 as a hardware support mechanism 12 to separate a plurality of OSs. Support machine Structure 12 is such that a Trust Zone (hereinafter referred to as TZ mechanism) can be used in a processor with ARM-based IP, and a VZ mechanism can be used in a processor with MIPS-based IP. VT mechanism can be used in the processor. The support organization 12 may also be a virtualized extension function, and a processor with an ARM-based IP may be exemplified by a Virtualization Extension and a LPAE (Large Physical Address Extension). For example, in the TZ mechanism 12, the memory space 12a of a non-secure environment can be separated from the memory space 12b of a closed environment, or the access to these memory spaces 12a and 12b can be separated, and a non-secure environment can be constructed. The protected environment is an environment in which the memory space 12b of the closed environment cannot be seen and cannot be accessed (access is prohibited).

Therefore, VAIOX20 functions as a sturdy firewall, and is designed to operate in a state where the conventional AI30 in a non-secure environment and the security AI50 in a closed environment are separated. On the one hand, in the state of overwriting of the memory space 12b that protects the closed environment, the security AI50 can perform input and output not only to the memory space 12b but also to the memory space 12a of the non-secure environment. Therefore, the security AI50 series can monitor the input and output of conventional AI30 and provide information to conventional AI30. The security AI50 in this example includes a monitoring function 59 that monitors the operation of the conventional AI30. The monitoring function 59 includes a function 58 for monitoring abnormal operation of the conventional AI 30. The conventional AI 30 includes a function of monitoring an abnormality of each functional module of the application group 47 operating in the non-secure environment 9. Therefore, by monitoring the abnormal operation of the conventional AI30, in a non-secure environment 9, it is possible to reliably monitor the cause of the function module for abnormal operation, and prevent the damage caused by it from happening to the minimum or to a minimum. . The occurrence of anomalies, such as known anomaly detection, predictive analysis, and omen monitoring, is not only by combining the anomalies pre-defined by these techniques and mechanical learning, but also by Probability of detecting unforeseen occurrences.

The hardware platform 10 includes a memory 13. The memory 13 includes a memory 13 a used in an open (normal, normal, non-secure, non-secure) environment, and a memory 13 b used only in a secured environment. The hardware platform 10 also includes various interfaces used in an open environment (Non-Secure World) 9, such as the use of visual / audible user interfaces 14, and the open connection to the Internet through various communication specifications. A network interface (first network interface) 15 of the network, and a connection interface 16 connected to an external device that can be used in an open environment.

The hardware platform 10 also includes various interfaces used in a closed environment (Secure World) 8, such as a network interface (second network interface) 17 for closed network connections in a home LAN, a car LAN, and the like, and A connection interface 18 for connecting external devices that can be used in a closed environment.

The super manager's VAIOX20 allocates the connection and control of the memory 13 and the interfaces 14 to 18 to the multifunctional OS21 and the security OS22, respectively. The open environment 9 and the closed environment 8 are built into the system 1. Conventional AI30 And the security AI50 series operate in an open environment and a closed environment, respectively.

Conventional AI30 is an AI that is generally used to deal with objects that have little damage even if they are captured (information). For example, the conventional AI30 recognizes the user's language or behavior, and when the recognized content is an Internet service, it is connected to the cloud side. The conventional AI30 functions as a support and consultant for ordinary Internet services, providing a large amount of information that is constantly changing, such as weather, diet, and information retrieval, and should be requested by users at different times. In addition, the conventional AI30 controls the function modules provided by third parties. For example, video content or services such as video-on-demand, and online shopping function modules can be cited.

On the one hand, the security AI50 series deals with objects that have been damaged as gateways (information). For example, the security AI50 is based on device control in the home (control by sensor 53 or switch 54), user personal data (Lifelog, such as action history, control device actions or user behavior, purchase history) 52 And personal services of user profiles. The security AI50 also includes the function of processing personal credit card information such as billing / shopping or the management and management of personal information (secret information) such as resume management, and encrypting it when submitting data to the conventional AI3055. In addition, the security AI50 includes a function 58 of monitoring the conventional AI30, and detects whether there is an unconditional attack on user actions and the like to determine whether it is different from the normal behavior to monitor whether the conventional AI30 is captured.

The Super Manager (VAIOX) 20 is used to communicate between the conventional AI30 and the security AI50, and guarantees that the data can be submitted only through the interface set by the developer. In this example, hardware virtualization technology is used to suppress improper access to the security AI50 side.

FIG. 2 shows a schematic configuration of a system including a hypervisor 20 including a virtualization support mechanism 20v that supports GPGPU (General Purpose Computing on Graphics processing units) as an example of the system 1. Assume that the processor unit 11 of the SoC (system-on-chip) with the in-vehicle equipment as the main applicable end includes: a plurality of CPUs (main processor units) 101 as a hardware platform; GPUs (graphics processor units) 102; and a computer resource 105 containing other functions (units) such as a network interface device controlled by them.

IVI (In Vehicle Infotainment) systems such as the navigation system 49b are often connected to a wide area network. (Registered trademark), Android (Rich OS) 21b to build a system. The graphic system 49c is a system that uses a lightweight and simplified RTOS (Real-Time OS) 21c certified for functional safety. RTOS21c is an OS (operating system) that is completely different from the multifunctional OS. Except for the CPU101, the graphics display part uses GPU102, etc., and is similar to the portrait-based app that operates on the multifunctional OS. . However, since the safety system 49c is important for functional safety, it will not be allowed to operate on the multifunctional OS.

Operation support systems such as ADAS (Advanced Driver Assistance System) 49d operate on the multifunctional OS21d. It is necessary to process a large amount of data input from an external camera in real time. In addition to the CPU101, in order to perform GPU 102 is used for image recognition calculation. The conventional AI30 also operates on the multi-functional OS21a. In addition to the CPU101, it is ideal for GPU102 for image recognition or high-speed processing.

For systems or function modules 30, 49b to 49d that operate on different OS21a to 21d, the hypervisor 20's virtualization support function 20v is created by virtualizing CPU101 and GPU102, and the resources containing CPU101 and GPU102 can be used in Each of the OSs 21a to 21d is used in a managed state, and a GPGPU from a plurality of applications can be used.

The hypervisor 20 includes a function for switching between the security mode 6 and the non-security mode 7. In the security mode 6, for this example, the CPU 101 is virtualized / para-virtualized as a resource allocation. Security AI50 or Security App 57 operates. In the security mode 6, not only the CPU 101 but also the GPU 102 may be allocated as a resource to the security OS 22.

FIG. 3 shows an example in which the system 1 having the conventional AI30 and the security AI50 is installed on a personal assistant device, such as a smartphone 71. The conventional AI30 recognizes the speech of the user 2 by using the voice recognition function 41b of the user interface, and provides a conversation with the user 2 by providing information to the user 2 through a voice synthesis / conversation agent 41a. The security AI50 monitors the content of the conversation through the VAIOX 20 through the monitoring function 59, monitors the actions of the conventional AI30, and autonomously judges whether the processing on the security AI50 side is necessary for the content of the conversation.

For example, when the user 2 decides to purchase the product "XX" in the conversation with the conventional AI30, the security AI50 recognizes that it will purchase the product "XX" according to the normal (intentional) judgment of the user 2 by monitoring its conversation . In the conventional AI30, the content of the voice-recognized conversation is converted into a text command and transferred to the store application 43a. The store application 43a uses the network connection function to access the cloud service 5 to purchase the product "XX". At that time, when credit card information, bitcoin information, or its authentication information is required for the store application 43a, in the security AI50, the payment function is stored from the secret information of the memory 13b stored in the security area managed by the personal data 52. 51 extracts the information required for purchase (information for checkout), and the encryption function 55 encrypts it and writes it into the memory 13a of the non-secure area accessible by the conventional AI 30.

The store application 43a or the conventional AI30 series does not need to be aware of the movement of the AI50. The store application 43a or the conventional AI30 is to use the encrypted checkout information prepared by the AI50 in the memory 13a, to check out the purchased product "XX", and write the information (purchase information) into the memory 13a. . Security AI50 needs to keep the record of purchase information in the security area, monitor the results of transactions between User 2 and conventional AI30, and judge that its actions are not inappropriate and the purchase information is safe. At all times, purchase information is fetched and added to the profile 52. Therefore, the conventional AI30 series can use the information that has been automatically prepared to enter the purchase process without realizing the existence of the security AI50.

FIG. 4 shows a case where the smart phone 71 is accessed improperly from the outside (outside). In a case where the conventional AI 30 did not notice the improper access invaded via the web service 43b, the improper access has tied the store application 43a. The security AI50 monitors the conversation with the user 2 and then monitors the communication content of the network service 43b, so that it is possible to recognize that improper access is occurring. Furthermore, when the security AI50 monitors the behavior of the store application 43a and does not agree with the conversation content of the user 2, it does not prepare information for checkout. Therefore, conventional AI30 series cannot advance to the purchase process because there is no information for checkout, and the user 2 can be notified of the abnormality. The security AI50 will not prepare information for checkout until the purchase between User 2 and regular AI30 is confirmed. By maintaining this action of the AI50, it is possible to prevent the conventional AI30 from unintentionally initiating actions due to improper access, and to prevent personal information such as card information from leaking.

FIG. 5 shows an example in which a system 1 including a conventional AI30 and a security AI50 is installed in a gateway system, such as a home gateway system (home gateway) 73. This home gateway 73 includes: a user interface 41 (41b) connected to a conventional AI (first artificial intelligence module) 30; and a home LAN connected to a security AI (second artificial intelligence module) 50 and / Or device control interface (switching function) 54.

In the home gateway 73, the conventional AI30 recognizes the utterance of the user 2 by using the voice recognition function 41b of the user interface, for example, saying "hope to be cool", and transforming the content of the utterance into a text command. The switch is written in the memory 13a of the non-secure area to be designated. Security AI50 series via VAIOX20 and The monitoring function 59 monitors the session. If the information written in the switching instruction of the memory 13a is consistent with the session, the switching function 54 is controlled according to the instruction to drive the programming application 43c in the non-secure area to perform appliances such as the air conditioner 4. Switch or adjust.

If the information of the switching instruction written in the memory 13a conflicts with the conversation, the security AI50 ignores the switching instruction written in the memory 13a and does not switch on or off appliances such as the air conditioner 4 without controlling the switching function 54 Adjustment. The security AI50 monitors the conversation with the user 2 to prevent improper use of the conventional AI30. In addition, the security AI50 can monitor past cases and actions outside the principle of human action for countermeasures. In addition, the security AI 50 may further include a function of turning off the switching function 54 to stop the operation of the home appliance when the setting or an abnormal command from the conventional AI 30 is repeated, until the user 2 instructs to perform the setting again.

Security AI50 uses VAIOX20 to use hardware virtualization technology, such as trusted areas, to block intrusions from conventional areas. Therefore, although the actions of the conventional AI30 will be monitored, the actions of the conventional AI30 will be monitored in the same manner as the conventional AI30 without being improperly invaded and affected by the abnormal actions of the conventional AI30, and will be implemented autonomously when the conventional AI30 is normal The intentional action of the conventional AI30, in the case of the abnormality of the conventional AI30, is to ignore the intention of the conventional AI30 or to self-discipline to normalize the abnormal intention of the conventional AI30.

There is a case where an instruction (operation) is input to the home gateway 73 from an external external terminal such as the smartphone 3. The security AI50 monitors the communication from the smartphone 3 through the network service 43b. If the communication is received through a secured path such as a VPN, it is recognized as a normal instruction and the smartphone 3 is permitted to operate Appliances. If the communication from the smartphone 3 is received via a non-secure path, it is blocked by ignoring the instructions from the smartphone 3 Improper access via home gateway 73.

An on-vehicle gateway or an IoT gateway having a configuration common to the configuration of the home gateway 73 can be provided. The on-board gateway has a user interface that is controlled or connected to the conventional AI30, and a vehicle LAN and / or device control interface that is controlled or connected to the security AI50. Conventional AI30 can also have a variety of interfaces including network connection.

The IoT gateway includes an Internet interface controlled or connected to a conventional AI30, and a factory LAN and / or device interface controlled or connected to a security AI50. It can be obtained from the factory LAN, etc. The information is provided to the collection center via a conventional AI30 and the Internet in a secure format. In addition, the security AI50 series can take only the information obtained from the collection center in a security format into a closed environment including the LAN in the factory.

The conventional AI30 and security AI50 are not limited as long as they are computer systems (software) that can learn, reason and judge. There are many types of artificial intelligence system.For example, it can also be data / knowledge AI (linear) including decision tree, inductive logic programming (ILP), reinforcement learning, Bayesian network (BN), support vector machine (SVM), etc. Type), can also be a brain-type AI including neural network (NL), deep neural network (DNN), or use fuzzy calculus, gene calculus, improved calculus, or use deep learning (Deep Learining ) Technology makes it wait for the innovators.

FIG. 6 shows an example of using a mechanical learning integrated platform (VAIOLIN platform, AI integrated platform) 80 provided by the applicant in the conventional AI30. In addition, the following so-called machine learning (ML) refers to those who include various technologies and architectures proposed or promoted to exert their functions as artificial intelligence. This AI integration platform 80 series can enable multiple MLs to facilitate the learning, inference, and judgment of an AI 30, and then, Contain and integrate different types of ML suitable for a wide range of problems.

The AI integration platform 80 includes: a plurality of mechanical learning modules 81a ~ 81d (to be represented as the ML module 81); a plurality of mechanical learning materials 82a ~ 82d used in each of the mechanical learning modules 81a ~ 81d (It should be set as the learning material 82 as the representative time); the mechanical learning module communication unit (VAIOX) 85 (20) that performs communication between multiple mechanical learning modules (ML modules) 81; suppresses multiple machines A mixed separation unit (VAIOLET) 83 of learning materials 82a-82d; a voice processing function 84 (41) of the user interface; a judgment unit 88 (38) for judging the failure and success of the user's response; The result of the determination unit 88 is an autonomous expansion unit 89 (39) that collects the mechanical learning module 81 and the mechanical learning data 82 from the network (open environment) and automatically constructs (automatically generates). The inter-module communication unit 85 (20) can also control the communication between the ML module 81 only in an open (non-secure) space. As described above, it can also separate the non-secure space from the secured space. The communication between the ML modules 81 is controlled in such a way that the non-secured space can be monitored from the preserved space.

For example, the mechanical learning module 81a includes various neural networks such as DNN, CNN, and GSN, and the mechanical learning module 81b includes boosting, DAE (Denoising Autoencoder), perception, etc. The learning module 81c includes SVM (Support Vector Machine), RF (random forest), and GMM (Gaussian Mixture Model, Gaussian Mixture Model), and the mechanical learning module 81d includes Bayes model (Bayes NP), multiplication and summing network , And DBN (Deep Belief Network). The techniques of mechanical learning are not limited by these.

In addition, the numbers in parentheses are shown in the system 1 with reference to FIG. 1. Number of units (modules) that share common functions. For example, the hypervisor 20 may also have the function of a communication unit 85 between mechanical learning modules that performs communication between a plurality of mechanical learning modules 81. Not only the CPU, GPU, and other computer resources, but also a plurality of mechanical The learning module 81 is virtualized and assigned to various functions provided by the AI integration platform 80.

Some examples of the functions provided by the AI unified platform (VIOLIN platform) 80 are: data exploration 87a with access analysis / data visualization / user mode construction; clustering / cold start of the attributes and characteristics of each data (0 data) Corresponding attribute data cluster 87b; network retrieval 87c including data submission to network data collection / ML and attribute data cluster; noise filter 87d for filtering noise generated on the network or speech and behavior learning; Task agent 87e performed by the user's task; memorizing the history and speech history of each user from the result of the task agent; 87f; forming a mixed pattern / vocabulary pattern for a new language and randomly generating a pattern / vocabulary feature analysis of the sentence 87g; Sensor I / F87h, which provides a physical interface for sensors, etc.

Figure 7 shows a map (configuration) for deep learning. Even if only the machine learning module (ML module) advocated by deep learning is illustrated, there are still deep networks suitable for learning, a complex neural network processing system, and a probabilistic model system suitable for processing a large amount of data. Wait. The AI integration platform 80 series may include one or a plurality of types of ML modules 81 shown in FIG. 7. In the AI integration platform 80, the communication between the various ML modules 81 can be integrated to function as one AI30 by the inter-ML communication unit 85 (20). Conversely, a learning unit 82 constructed by separating each ML module 81 by a separation unit (VAIOLET) 83 is prepared to prevent damage to the learning materials (contami), so that a plurality of ML modules 81 can perform desired functions. Therefore, it is possible to combine conventional routines suitable for personal assistance systems 71, various gateway systems 73, and the like. AI30 is the best solution (ML module) 81. The AI integration platform 80 can also be used to construct the security AI50.

This AI integration platform 80 includes a success judgment unit 88 (38) for judging the failure and success of the response to the user 2 and an autonomous expansion unit 89 (39) for adding a new ML module 81 as needed, which can be accelerated by failure. personalise.

As shown in FIG. 8, the AI integration platform 80 performs dynamic learning 92 including individual learning from the initial learning 91 based on teacher data, which is suitable for providing improved AI 30. For example, the AI30 equipped with the AI integration platform 80 is started with the smallest ML module 81 at the beginning of the initial learning 91 (0 歳), and obtains basic intention estimation, prediction support, and task agency through education using teacher materials. , The function of self-discipline learning. After that, in the dynamic learning 92, based on the actions performed under the user 2, the success / failure judgment unit 88 judges the failure and success of the response to the user 2. With the failure, the autonomous expansion unit 89 adds a new ML module 81 as necessary. It can speed up personalization. In particular, it is not limited to the characteristics of the initial ML module 81. By adding an ML module 81 having different characteristics, it is possible to provide a more flexible and personalized AI30.

FIG. 9 shows an example of an on-vehicle gateway 75 on which the system 1 including AI 30 and 50 is mounted. The on-vehicle gateway 75 includes a function of automatically determining and further determining the correct answer of the user's viewpoint. The success / failure judgment unit 88 of the AI integration platform 80 documents failures and correct answers and makes a judgment, and the self-discipline expansion unit 89 reflects the results to the system 1, which greatly improves the basic performance of the AI 30. In particular, this system 1 is a chip-mounted system. By combining the ML module 81 and the learning materials 82 on the AI integration platform 80, the performance of the system 1 in the local environment alone can greatly improve the performance of AI30. Therefore, in the local environment where the information is individually held and the center where the information is centralized, the user's information can be collected individually. In some cases, the user 2 can be reliably assisted or the in-vehicle device can be normally operated.

FIG. 10 shows the basic mode construction based on collective intelligence in the initial learning 9 and the recommendation based on the personalized feature amount in the dynamic learning 92. In the initial learning 91, a basic mode construction based on collective intelligence, which is a basic structure of the AI integration platform 80, and a basic mode of feature space are generated. In the dynamic learning 92, according to personalization, the mapping of the individual characteristic quantities and the update of the type selection mode are performed on the AI integration platform 80, and various recommendations can be made according to personal preferences and actions.

FIG. 11 shows the process of defining personal characteristics in the dynamic learning 92 from the huge amount of data in the initial learning 91. In this AI integration platform 80, learning materials 82 can be locally generated and local performance can be exerted. At the same time, the combination with the server can also be used for composite correspondence. In the initial learning 91, access to the huge amount of data stored on the server side can be performed, and individuals can be extracted from the huge amount of data in the dynamic learning 92. Information to obtain personal characteristics. In addition, it is also possible to change the thinking of AI30 from generalization to general assistance to avoid danger and so on. For example, in an on-vehicle gateway, a personalized dangerous operation can be changed to general assistance to avoid danger, or based on dangerous information such as navigation to general assistance.

FIG. 12 shows that the on-vehicle gateway 75 of the system 1 including the conventional AI30 and the security AI50 borrows a voice dialogue platform 41c of the voice processing function 41 and talks with the user 2 while learning the user's actions, and autonomously expands the AI integration platform 80. In addition, in a manner that improper behavior cannot be performed, the security AI50 monitors the processing of the conventional AI30.

This dialogue platform 41c is a voice dialogue platform for the purpose of collecting information of the user 2 speaking in the car, collecting the content of the user 2 speaking in the car, and The center has a huge amount of information. The dialogue platform 41c is a platform that reduces the number of times of talking with the user 2 and provides a question and answer in a manner that can provide a purpose in a shorter time to realize a natural dialogue with the user. Regarding the content contained in the dialogue (for example: reservation, shopping, information guidance, etc.), the task that meets the needs of users is directed to the dialogue. And natural conversations can be achieved by blending gossip conversations with task-oriented conversations.

At this time, the success / failure determination unit 88 is to successfully guide the dialogue for the user 2. When it is determined that the ML module 81 mounted on the current AI integration platform 80 is insufficient, the self-discipline expansion unit 89 may also be downloaded from the Internet ( Cloud) and other open environments, automatically collect the ML module 81 and the learning materials 82 as the basis for successfully guiding the dialogue, and install them on the AI integration platform 80. For example, if the ML module 81 that can respond to music in the area of interest expressed by the user 2 is not installed in the initial learning 91, the dynamic learning 92 may be installed. It is also possible to create a way to be installed on the AI integration platform 80, and set the storage of the ML module 81 suitable for various fields of response and the learning materials 82 for the initial learning of the ML module 81 in the cloud. System 1 The ML module 81 can be expanded by accessing the memory as needed.

The dialogue platform 41c is to respond to a wide range of interests that the user cares about by means of a chat, and to turn into a specific task-oriented dialogue such as shopping. Compared to induction, providing natural recommendations for a wide range of topics is to be expected.

Security AI50 monitors these sessions using conventional AI30. As a result of session induction, when shopping, the security AI50 recognizes it and prepares the necessary checkout information for shopping in a secured state, such as encryption, so that the conventional AI30 can be used for shopping.

The change or addition of the ML module 81 is also effective in improving the function of voice recognition. For example, as shown in FIG. 13, for input sound, voice recognition is being used In the case of a low task success rate, the intent understanding engine 81x initially installed on the rule base of the AI integration platform 80 is changed to a statistical intent understanding engine 81y by the success / failure determination unit 88 and the self-discipline expansion unit 89, even though the voice recognition engine 81v The same rate of correct identification can also improve the success rate of the task.

Although the above uses several examples of the system 1 containing conventional AI30 and security AI50 to illustrate the present invention, these are only examples, and the present invention is not limited by these, and can be preferably used in the application of artificial intelligence ( AI) or services provided by all areas, function modules, devices, terminals, and networks with the possibility of AI support.

1‧‧‧ system

6‧‧‧ Security Mode

7‧‧‧ non-security mode

8‧‧‧ closed environment

9‧‧‧ Non-secure environment

10‧‧‧hardware platform

11‧‧‧ processor unit

12‧‧‧ Support Agency

12a, 12b‧‧‧Memory space

13, 13a, 13b‧‧‧Memory

14‧‧‧user interface

15, 17‧‧‧ web interface

16, 18‧‧‧ connection interface

20‧‧‧ Hyper Manager

21‧‧‧ Non-Security OS

22‧‧‧Security OS

30‧‧‧ conventional AI

38‧‧‧ Response Failure and Success Unit

39‧‧‧ Unit of Personal Database

41 ~ 44, 51 ~ 55‧‧‧Function modules

50‧‧‧Security AI

57‧‧‧Security App

58‧‧‧Function of abnormal movement

59‧‧‧Monitoring function

Claims (15)

A system containing artificial intelligence, comprising: a first artificial intelligence module including a function of assisting or autonomously performing information exchange in an open environment; and a second artificial intelligence module that monitors the aforementioned first artificial intelligence module The operation voluntarily performs the process accompanied by the aforementioned information exchange in a closed environment. For example, the artificial intelligence-containing system of claim 1, wherein the aforementioned open environment includes a user interface, a first network connection interface for connecting to an open network service, and stored in the aforementioned user interface or the first network The first memory for information exchanged between the connection interfaces. The closed environment includes the second memory that cannot be accessed by the first artificial intelligence module. The second artificial intelligence module contains information that will accompany the foregoing information. The information requested and exchanged is the function of encrypting the secret information stored in the aforementioned second memory to the aforementioned first memory. For example, the artificial intelligence-containing system of claim 2, wherein the closed environment includes a second network connection interface for connecting to a closed network service. For example, the artificial intelligence-containing system of any one of claims 1 to 3, wherein the second artificial intelligence module includes a function of monitoring the abnormality of the first artificial intelligence module. For example, the artificial intelligence-containing system of any one of claims 1 to 4, which includes: the first OS supporting the aforementioned open environment and the first OS operating the aforementioned first artificial intelligence module; and the first OS supporting the aforementioned closed environment. 2OS is also the second OS operated by the aforementioned second artificial intelligence module. For example, the artificial intelligence-containing system of claim 5, which includes: a processor unit including: a memory containing a secured area and a non-secured area; and a support organization forbidding access to the aforementioned secured area; and a hypervisor, in the aforementioned Operated on a processor unit, the hypervisor includes: a mechanism that permits the second OS to access the secure area and the non-secure area and prohibits access to the open environment; and permits the first OS to the non-secure area. And the access to the aforementioned non-secure device, and the use of the aforementioned support mechanism to access to the aforementioned security area is prohibited. If the artificial intelligence-containing system of claim 6, wherein the aforementioned processor unit includes: at least one main processor unit; and at least one graphics processor unit, the hypervisor includes: the at least one main processor unit and The virtualization support function of the at least one graphics processor unit. For example, a system containing artificial intelligence according to any one of claims 1 to 7, wherein the aforementioned first artificial intelligence module includes: a plurality of mechanical learning modules; a plurality of mechanical learning materials used in each mechanical learning module; and performing the foregoing Multiple machinery A communication unit between mechanical learning modules for communication between the learning modules; and a separate unit that suppresses the mixing of the aforementioned plurality of mechanical learning materials. For example, the artificial intelligence-containing system of claim 8, wherein the aforementioned first artificial intelligence module includes: a unit for judging the failure and success of the response to the user; and automatically collecting the response for the user from the aforementioned open environment The successfully guided mechanical learning module and mechanical learning data are installed in the installation unit of the aforementioned first artificial intelligence module. For example, the artificial intelligence-containing system of any one of claims 1 to 8, wherein the aforementioned first artificial intelligence module includes: a unit for judging the failure and success of the response to the user; and automatically collecting and using the information from the aforementioned open environment for A unit of a personal database will be automatically constructed for the information successfully guided by the user's response. The artificial intelligence-containing system according to any one of claims 1 to 10, which includes a control module equipped with the aforementioned first artificial intelligence module and the aforementioned second artificial intelligence module. The artificial intelligence-containing system according to any one of claims 1 to 11, wherein the system is a personal assistance system including a user interface connected to the aforementioned first artificial intelligence module. The artificial intelligence-containing system according to any one of claims 1 to 11, wherein the system is a home gateway system, which includes a user interface connected to the aforementioned first artificial intelligence module and an artificial intelligence connected to the aforementioned second artificial intelligence. Module's home LAN and / or device control interface. The artificial intelligence-containing system according to any one of claims 1 to 11, wherein the system is an on-vehicle gateway system, which includes a user interface connected to the aforementioned first artificial intelligence module and connected to the aforementioned second artificial intelligence. Module's on-board LAN and / or device control interface. The artificial intelligence-containing system according to any one of claims 1 to 11, wherein the system is an IoT gateway, which includes an Internet interface connected to the aforementioned first artificial intelligence module and an artificial interface connected to the aforementioned second artificial intelligence. Factory LAN and / or device interface of the module.