迈向通用人工智能：原理、技术与路径探讨

版权申诉

32 浏览量更新于2024-06-14 收藏 1.06MB PDF 举报

"《通用人工智能：原因、内容与方式》一文探讨了当前人工智能芯片发展的重要性和方向。随着智能化感知、自动化以及边缘计算应用市场的推动，AI芯片的设计越来越关注实现低功耗和低成本的神经计算。为了将应用特定的AI芯片过渡到具有人类类似智能能力的通用人工智能（AGI），文章提出了跨学科研究领域的一次全面概述。首先，作者强调了向AGI转变时所需考虑的关键因素，包括但不限于算法的普适性、性能提升、鲁棒性增强以及可扩展性。这涉及到如何在硬件层面模仿和提升人类智慧，构建能够处理多种任务并具备学习和适应性的系统。文章进一步列举了新兴的AI芯片技术，如深度学习加速器、神经形态芯片等，并对边缘AI的不同实施策略进行了分类，比如云计算、物联网(IoT)环境下的本地处理等。设计流程方面，本文详述了从需求分析、架构设计、原型开发到测试验证的完整路径，强调了设计过程中对神经网络模型优化、数据处理效率和能效管理的重要性。最后，作者列出了构建AGI芯片所需的关键设计要素，包括硬件架构的灵活性、高效的数据流管理和高级别的并行处理能力。同时，文中还讨论了如何通过模拟人类认知过程、评估指标和基准测试来验证AGI芯片的性能和有效性。《通用人工智能：原因、内容与方式》提供了深入理解AGI芯片开发挑战与机遇的视角，对于AI工程师、硬件设计师和研究人员来说，这篇论文是一份宝贵的参考资料，帮助他们探索如何在技术进步中迈向更加智能且普遍适用的AI系统。"

of Tabula Rasa [23], where the mind is considered as a blank

slate, and anything can be written on it that is enriched by

experiences in time. An untrained neural network can be

considered as a ’Tabula Rasa’ in the sense that it can be trained

to perform a given task through weight updating algorithms.

Although, it may make sense to establish this correlation, it

is an incorrect assumption given the various progress in the

developmental studies related to intelligence contrast this.

3) Human intelligence: Several studies from developmental

psychology point out that both evolution and blank slate only

approach to intelligence is wrong. Instead, the mind is capable

of high degree of generality, not limited by innate skills and

capable to acquire new skills throughout the life. At the same

time, our cognitive abilities are specialised by evolution that

is driven by biological priors and evolutionary limitations that

dismantle ’blank slate’ theories as well.

The fact that humans are capable of excelling in speciﬁc

tasks and skills compared to many other animals are a

testament to this. Understanding human cognitive priors are

essential to developing human like intelligent forms.

There are several priors that help humans perform intelligent

tasks [24]. They can be grouped as:

1) Motor-sensory priors: These are motor-sensory skills

that humans are born with such as ability to move,

react and respond to sensory inputs. The reﬂexes such as

vestibulo-ocular reﬂex is a good example of this, where

the head movements coordinate strongly with the eye

movements.

2) Meta-learning priors: These are the underlying strategies

what deﬁne how we learn any tasks. For example, the

idea of functional modularity, hierarchical information

processing, spatio-temporal information coding, decod-

ing and organisation can be classiﬁed under this.

3) Knowledge priors: These priors gives a broader sense

of the environment around us. For example, the shapes

of objects, innate indication about space and depth,

innateness of numbers, sense of time, and sense of social

intuitions.

The development of AI hardware that cater to only motor-

sensory priors such as using sensors and responses helps to

imitate an artiﬁcial human body without general intelligence

abilities. While, meta-learning priors form the key aspects of

intelligence that enable learning and help solve problems, AI

chips today make use of various forms of meta-learning priors,

and are in constant effort to identifying newer and efﬁcient

learning implementations. In contrast, the knowledge priors

help assess the AI chip implementations against human like

intelligent tasks. The AI chips that could potentially include

programmed number of knowledge priors could trick the

speciﬁc test to show human like intelligence ability, however,

they do not represent general intelligence, and such system

should not considered as AGI systems.

C. From General Intelligence to AGI

AGI aims to embed general intelligence abilities in ma-

chines and are also known as ’full AI’ or ’strong AI’ systems

[4], [5], [25]. There is no evidence of practical strong AI

systems that exist today that matches fully with human like

intelligence, although there have been active attempts to build

libraries and software (e.g. see [26]) to build such systems.

While, these software and libraries are essential to progress

the AGI, they alone cannot achieve physical compactness

and energy efﬁciency required for such complex systems.

This is where, AGI chip development becomes important,

where either they become supportive tool as an accelerator

to algorithms or that can have innate priors embedded within

its hardware.

The knowledge priors are extensively used in testing AGI

algorithms, by giving and comparing human-like challenges to

AGI system. Examples include: (1) testing the AGI system by

making it take a university program and obtaining a degree,

(2) testing the ability of the AGI to perform in a job, (3) testing

the ability of AGI system to make a coffee, and (4) perform

Turing test [5], [27]–[30].

The AGI system should be able to solve ’AI-hard’ prob-

lems [31], [32], that would require the use of natural lan-

guage processing, computer vision, and understand images,

emotions, and environment. With the current state of the art AI

system, solving an AI-hard problem requires the assistance of

humans in addition to the acceleration provided by computing

machines.

Majority of AI methods aim to mimic the functional and/or

structural behavior of the neural networks. The functional

understanding of brain is emulated through knowledge repre-

sentation, statistical models and complex networks. Mimicking

biology driven human intelligence will require the application

of the concepts of feedback and feedforward propagation of

information, and integration of sensing mechanisms, along

with meta-learning priors and knowledge functions speciﬁc

to tasks.

The brain is considered as the benchmark system for AGI re-

search. There are many approaches in neuromorphic computa-

tion that aim for achieving energy efﬁciency and performance

benchmarks of the brain [33], [34]. The performance of the

AI systems has considerably improved due to the availability

of high-performance computers and excessive capturing of

labeled data. As more and more sensors are connected to the

devices, the volume, veracity, and velocity of data increase.

The deep learning architectures [35] can use the data to arrive

at improved classiﬁcation and recognition accuracy.

D. AI Hardware

The neuromorphic computing is inspired by the information

processing in a human brain by mimicking the operation

principle and structure of the neurons and synapses (Figure 1

(a)). Low power intelligent information processing is a key

concept in neuromorphic computing inspired by the human

brain which performs complex processing tasks consuming

nearly 20 W of power. The ﬁrst task-speciﬁc neuromorphic

systems have been developed for speciﬁc applications, such as

object recognition or prediction of speciﬁc data. The second

generation of the neuromorphic systems is moving towards the

tasks corresponding to human cognition, like adaptation and

interpretation of the unknown information, and a ﬁnal goal of

building strong AI systems [36], [37].

剩余14页未读，继续阅读

百态老人

粉丝: 9848
资源: 2万+

迈向通用人工智能：原理、技术与路径探讨

人工智能的主要内容与及方法.pdf

通用AI，通用技术，通向何方.pdf

通用人工智能的问题和机遇.pdf

通用人工智能的危险和前景.pdf

通用人工智能与人类心智模型.pdf

通用人工智能的执行入门书.pdf

智能治理：通用人工智能时代的治理命题.pdf

简介：通用人工智能的各个方面.pdf

AI大模型赋能人形机器人，迈向通用人工智能的一大步.pdf

Is Artificial General Intelligence Impossible? 通用人工智能不可能吗？.pdf

最新资源