GB/Z 177.1-2026Intelligence grading of artificial intelligence terminal - Part 1: Reference framework (English PDF)

Foreword

This document was drafted in accordance with the rules given in GB/T 1.1-2020, Directives for standardization - Part 1: Rules for the structure and drafting of standardizing documents.

This document is Part 1 of GB/Z 177, Grading of intelligence for artificial intelligence terminals.

Attention is drawn to the possibility that some of the elements of this document may be the subject of patent rights. The issuing body of this document is not to be held responsible for identifying any or all such patent rights.

This document was proposed by and is under the jurisdiction of the National Information Technology Standardization Technical Committee (SAC/TC 28).

The drafting organizations include the China Electronics Standardization Institute; the China Academy of Information and Communications Technology; the China Center for Information Industry Development; Xiaomi Communications Co., Ltd.; Honor Device Co., Ltd.; Huawei Technologies Co., Ltd.; OPPO Guangdong Mobile Telecommunications Co., Ltd.; Lenovo (Beijing) Co., Ltd.; Vivo Mobile Communication Co., Ltd.; and iFLYTEK Co., Ltd., among others. Twenty-nine drafters are named.

1 Scope

This document provides a reference framework for artificial intelligence terminals and sets out the elements of their intelligence capability.

This document is intended to guide the grading of intelligence for artificial intelligence terminals of every kind, and provides a reference for their design, development, application, selection and evaluation.

2 Normative references

The following documents contain provisions which, through normative reference in this text, constitute indispensable provisions of this document. For dated references, only the edition cited applies. For undated references, the latest edition of the referenced document, including any amendments, applies.

GB/T 32400, Information technology - Cloud computing - Overview and vocabulary

GB/T 41867, Information technology - Artificial intelligence - Terminology

3 Terms and definitions

For the purposes of this document, the terms and definitions given in GB/T 32400 and GB/T 41867 and the following apply.

3.1 artificial intelligence terminal - a terminal product that has the capability of active perception and understanding, multimodal interaction, intelligent services and learning and evolution, and that completes specific tasks. Note 1: the flow of an intelligent task generally involves perception, planning, decision, execution and learning. Note 2: an artificial intelligence terminal is made up of software and hardware; the software includes the artificial intelligence model, the intelligent applications, the operating system, the user interface and the device-cloud collaboration interface, and the hardware includes the communication module, the processor, the internal storage, the peripheral input and output devices and the display.

3.2 user - the person who uses the artificial intelligence terminal. Note: in some interaction scenarios the user is an intelligent system, such as an artificial intelligence terminal or an agent.

3.3 multimodal interaction - a mode of interaction in which information is input and output between the user and the terminal through several means of communication. Note: those means include speech, text, images, gestures, touch, eye movement and expression.

3.4 context awareness - acquiring, understanding and using information about the user, the environment, the task and the state of the device itself.

3.5 intent understanding - recognizing the underlying goal or need from the input of the user, combined with context awareness.

5 Reference framework

5.1 Overview. An artificial intelligence terminal combines the artificial intelligence capability of the device side and of the cloud side, and comprises the hardware devices, the operating system, the typical applications, the intelligence modules and the security management, as shown in Figure 1. The terminal receives the input of the user, perceives the operating environment, understands the instructions and the operating intent of the user and carries them out; where necessary it connects over the network to the AI cloud and to information on the internet for collaborative enhancement, and it connects to external devices for device control and interaction. Depending on the class of terminal (see Clause 6) the modules are tailored, and the capability is delivered on the device side, on the cloud side or by a device-cloud mix.

5.2 Hardware devices provide the computing, storage, interaction and interconnection the terminal needs, and are the foundation of the artificial intelligence terminal: the computing unit, the storage unit, the interaction unit and the communication unit.

a) The computing unit is responsible for general computing and for AI computing, and includes the general-purpose processor (CPU), units designed specifically for neural network computing (NPU), digital signal processors for audio and images (DSP), graphics processors for general parallel computing (GPGPU) and accelerated processors combining CPU and GPU (APU). b) The storage unit carries the operating system, the applications, the AI models and the user data, and includes chip flash such as eMMC and UFS, solid-state disks and memory (RAM).

c) The interaction unit is the entry point through which the terminal perceives the physical world and acquires data, and includes but is not limited to: standard input devices such as keyboard and mouse; visual sensors such as CMOS image sensors (CIS) and infrared thermal imaging sensors; audio sensors such as microphone arrays; motion sensors such as accelerometers, gyroscopes, magnetometers and vibration motors; biological sensors such as heart rate, blood oxygen saturation and body temperature sensors; and environmental sensors such as light, distance, barometric and temperature and humidity sensors. d) The communication unit is the key to connecting the terminal to the network, to device-cloud collaboration and to device interconnection, and includes but is not limited to network links such as wired networks, wireless LAN and mobile networks (4G, 5G), and short-range wireless communication such as Bluetooth, NearLink, NFC and ZigBee.

6 Classification of terminals

6.1 Overview. This document gives a classification reference along the dimensions of computing power, scenario of use, number of users, AI operating mode, interaction mode and manner of use; the class to which each kind of terminal product belongs is given in Annex A.

6.2 Computing power. By the strength of their computing power, artificial intelligence terminals are divided into: a) devices of strong AI computing power, which integrate dedicated and extensible processor units such as GPU, NPU and APU and can complete complex AI tasks independently or mainly on the device side, such as running medium and large local models, real-time high-definition image or video analysis and complex multimodal interaction (Note 1: generally running large language models of not fewer than one billion parameters); b) devices of weak AI computing power, which have some local AI computing power, achieve limited acceleration through the CPU and the DSP and can handle lightweight AI tasks such as keyword wake-up and simple voice command recognition, relying on cloud-side collaboration for complex tasks (Note 2: they generally run AI tasks while connected to the network).

6.3 Scenario of use. Terminals are divided into: a) personal terminals, highly portable, highly personalized (such as personalized product recommendation), multifunctional and attentive to privacy protection; b) office terminals, which raise the efficiency of collaboration (such as AI-assisted programming), process information (such as intelligent meeting minutes) and serve professional applications; c) home terminals, shared by several users, for control of the home environment, content services, far-field speech interaction, multi-user recognition, linkage of home devices and content recommendation; d) transport terminals, which provide travel services, navigation optimization, perception of the in-vehicle environment, fatigue monitoring and multimodal interaction in the cabin space of the vehicle; e) learning devices, aimed at education, providing personalized teaching support such as intelligent tutoring, with functions such as analysis of knowledge mastery and reinforcement practice on weak points, for example language learning practice.

6.4 Number of users. Terminals are divided into: a) single-user terminal devices, used mainly by one user, without restricting login by several user IDs, such as mobile terminals and mini computers; b) multi-user terminal devices, which can recognize the identity of the user automatically during AI interaction and switch usage habits, such as smart speakers.

6.5 AI operating mode. Terminals are divided into: a) pure device-side operation, where the AI model and data processing are performed mainly or entirely locally on the terminal, with no or very little reliance on a cloud connection (Note 1: low latency, high privacy and available offline; limited by the computing power and storage of the terminal, the scale and complexity of the model are limited); b) device-cloud collaborative operation, where the terminal and the cloud-side AI capability work together and the computing load is allocated dynamically according to the nature of the task, the state of resources and privacy needs (Note 2: it combines the real-time response and privacy of the device side with the strong computing power and large data of the cloud); c) pure cloud-side operation, where the terminal serves mainly as the data acquisition and interaction interface and AI computing and decision are performed entirely in the cloud (Note 3: low demand on terminal computing power, using powerful cloud models and data, dependent on the network connection, higher latency and comparatively greater privacy risk).

7 Capability elements

7.1 Perception capability covers, among others, perception of software information such as resource occupancy, installed software and running state, and 7.1.5 perception of interconnected devices, which includes but is not limited to: a) perception of externally connected devices, recognizing the connection state and the capability of the device; b) collaborative perception, sensing useful information through interconnected devices, for instance sensing the movement and health state of the user through a wearable device.

7.2 Cognitive capability. 7.2.1 Definition: the capability of the terminal to understand, reason about, plan and reflect on information on the basis of perceived data, through logical rules, knowledge graphs and machine learning models, so that the execution of the task matches what the user expects.

7.2.2 Intent understanding: the capability to recognize the underlying goal or need from the input of the user by combining context awareness of every kind. a) Active prediction and passive execution: active prediction of intent, proposing a suggestion or an action without an instruction from the user, such as proposing a change of travel plan on a cold wave warning or recommending a flight from the diary; passive execution of intent, carried out only after an instruction from the user, such as asking it to correct an essay. b) Single-modal and cross-modal input understanding: a spoken query such as what is the French translation of apple; or a cross-modal query such as pointing with a gesture and asking what is this or look this word up for me. c) Single-step and multi-step intent: a single-step intent achieved in one exchange, such as setting an alarm for six tomorrow morning; a multi-step intent needing several exchanges, such as booking the early flight to Beijing next Saturday. d) Explicit and implicit intent: explicit, such as open the recorder; implicit, such as the room is too dark.

7.2.3 Clarification by follow-up question: where the input of the user is insufficient, such as lacking enough context or personalized data, is vague, such as dialect or background noise, or is ambiguous, so that the system cannot understand the intent or plan the task accurately, the terminal asks of its own accord, including but not limited to: a) before the task is carried out, asking the user for details of what the task requires; b) after the task is carried out, offering several possible supplementary execution options.

7.2.4 Reasoning capability: the capability to process explicit information and rules in depth through logical deduction or contextual reasoning on the basis of the information perceived, and so to generate new conclusions or insight, including but not limited to: a) common-sense reasoning, including fact verification, conditional reasoning, similarity judgement and common-sense conclusions; b) scientific reasoning, in which the terminal recognizes the science implicit in the question of the user and solves the problem using mathematical concepts and scientific principles; c) spatio-temporal reasoning, in which the terminal analyses time and space information collected from the input of the user or from sensors.