Bridging the Gap between Artificial Intelligence and Artificial General Intelligence A Ten Commandment Framework for Human -Like Intelligence

2025-04-30 0 0 891KB 51 页 10玖币

侵权投诉

Bridging the Gap between Artificial Intelligence and Artificial

General Intelligence: A Ten Commandment Framework for

Human-Like Intelligence

Ananta Nair1,2 and Farnoush Banaei-Kashani1

1. University of Colorado, Denver, 2. Dell Technologies Inc

Abstract

The field of artificial intelligence has seen explosive growth and exponential success. The last

phase of development showcased deep learnings ability to solve a variety of difficult problems

across a multitude of domains. Many of these networks met and exceeded human benchmarks

by becoming experts in the domains in which they are trained. Though the successes of artificial

intelligence have begun to overshadow its failures, there is still much that separates current

artificial intelligence tools from becoming the exceptional general learners that humans are. In this

paper, we identify the ten commandments upon which human intelligence is systematically and

hierarchically built. We believe these commandments work collectively to serve as the essential

ingredients that lead to the emergence of higher-order cognition and intelligence. This paper

discusses a computational framework that could house these ten commandments, and suggests

new architectural modifications that could lead to the development of smarter, more explainable,

and generalizable artificial systems inspired by a neuromorphic approach.

Introduction

Though the concept of artificial intelligence (AI) may seem like a futuristic prospect, the desire to

create man-made intelligence has long been a human yearing. Dating back to approximately 700

BCE is the ancient Greek myth of Talos, the first robot. Talos was created by the god Hephaestus

to serve as guard of Crete and hurl boulders at incoming vessels. One day a ship approached the

island, which, unknowing to the behemoth automatron, would serve as his greatest challenge. In

an attempt to escape the machine, Medea, a sorceress onboard, constructed an ingenious plan

and offered the robot a bargain; granting him eternal life in return for removing his only bolt.

Surprisingly, this offer resonated with Talos, who had not yet come to grips with his own nature,

nor understood his longings for human desires such as immortality. Though the tale ended

tragically for the giant, it does speak to the desires and fears humans have long had of creating

intelligence and the blurred line between man and machine, a challenge that is becoming

increasingly prominent today.

From mechanical toys of the ancient world to the looming dystopian apocalyptic scenarios of

science fiction, the creation and progression of AI has long been on the human mind. Though

many would argue the field has seen many bursts and busts, the last decade has resulted in the

most remarkable progress till date. Largely these advances have come from deep neural

networks that have become experts in areas such as vision, natural language processing, and

reinforcement learning (Brown et al., 2020, Ramesh et al., 2022, Chowdhery et al., 2022,

Schrittwieser et al., 2020, Ye et al., 2021, Baker et al., 2019, Arnab et al., 2021). Their breadth of

application is expansive and the commercial and practical deployments being undertaken seem

almost endless. Applications such as facial recognition ((Balaban et al., 2015), have found

common place on our phones and in our daily lives; virtual assistants like Alexa are a significant

step up from IBM’s Shoebox (Soofastaei, 2021), that could recognize only sixteen words and

digits; online translators are capable of accurately translating between any two or multiple

languages (Fan et al., 2021), and reinforcement learning agents have achieved human and

superhuman performance on a range of complex tasks such as board and video games

(Schrittwieser et al., 2020, Silver et al., 2018, Berner et al., 2019, Vinyals et al., 2019). Even in

academia, these tools have resulted in significant breakthroughs in a diverse set of endeavors,

including weather prediction, protein unfolding, mental health, medicine, robotics, and

astrophysics (Ravuri et al., 2021, AlQuraishi, 2019, Su et. al, 2020, Lusk et al., 2021, Gillenwater

et al., 2021, Pierson & Gashler, 2017, Huerta et al., 2019, George & Huerta, 2018).

Inarguably, each of these undertakings deserve to be commended for its unique

accomplishments. However, there still remains much work to be done by the field to achieve

general intelligence akin to the natural world. Deep networks in collaboration with the

advancement of GPUs, have accelerated data-processing to master pattern recognition and other

statistical based learning, supervised and rule based learning, unsupervised and training-test

generalization, and reinforcement and multi-agent learning (Zhu, 2005, Caruana & Niculescu-

Mizil, 2006, Sutton & Barto, 2018, O'Reilly et al., 2021, Mollick et al., 2020, Baker et al., 2019,

Stooke et al., 2021). These ingenious mathematical algorithms and data-intensive processing

techniques lead to great success in ideal conditions by converting a sparse problem space into a

dense sampling whose distribution can be integrated over or overfit on. These networks can even

further transfer to out-of-distribution learning, doing so by integrating over both training and testing

sets (Vogelstein et al., 2022). However, a common side-effect and growing problem of this type

of approach has been the creation of data-sensitive black boxes that struggle to generalize

outside of their well defined parameterizations (Lake et al., 2017, Geirhos et al., 2020). These

challenges make current AI tools great optimizers but they still fall short of traditional intelligence.

In comparison, natural intelligence is an exceptional example of evolutionary engineering, with

the human brain regarded as the premier example. Given the latest successes but large

limitations of deep networks, the field of AI is now more than ever drawing comparisons and

parallels between what neural networks can do and what natural intelligence is capable of

(Vogelstein et al., 2022, Silver et al., 2021, Richards et al., 2019). The brain, unlike AI, does not

demonstrate or strive for exceptional performance at every task it attempts in a hopeless effort to

maximize its reward function. Instead, it creates exceptional universal learners that are capable

of generalizing their abstract representations and skills to learn any task quickly and easily. This

methodology does not lead to exceptional general performance across all tasks, but instead

behavior is adjusted through goal-driven learning to create a system that only excels at high value

objectives whereas other tasks have more moderate probabilities of success. This type of learning

optimizes for specific performance, while leaving enough computational resources to adequately

perform all tasks. For example: athletes and soldiers may excel at tasks that require scene

integration, strategy, and physical or reflex capabilities, however scientists would not need to

excel in these domains and may instead showcase exceptional performance on logic and

reasoning, or artists on fine-tuned motor movements. Though each profession can learn the

other's skill set, humans prioritize their performance based on high value objectives that align with

their goals.

We believe this methodology of maximizing an architectures learning based on creating general

goal-directed learners rather than overtrained task specific agents is the key to the emergence of

more intelligent systems. In this paper, we use the brain as inspiration, to identify the key

properties that we believe are essential to higher order cognition. The next section of the

introduction teases apart components of brain function that could address notable limitations seen

in current AI tools. Whereas, the last section of the introduction provides an overarching summary

of the framework. Next, section two addresses the key properties of brain function that we believe

lead to intelligence. These are communicated as the ten commandments, and we believe

intelligence emerges not from an individual commandment but rather an all encompassing system

whose collective sum is stronger than its part. Lastly, section three addresses how these

commandments can be proposed into a framework for the development of AI systems that

transition from statistical superiority to general intelligence. Though this paper is largely

addressing the assembly of components that may lead to autonomous AI akin to natural

intelligence, we believe that the individual commandments in themselves can be beneficial to

improve a myriad of AI systems and tools where the end goal is not an all encompassing system.

Through the presentation of the commandments and the framework, we believe models can be

scaled up or down as needed.

1.1 Teasing the Brain Apart; The Paradox of Intelligence

Numerous deep learning models have surpassed human performance on tasks on which they

have been intensively trained (Schrittwieser et al., 2020, Silver et al., 2018, Berner et al., 2019,

Vinyals et al., 2019). This has led some to even argue that these models are akin to humans in

the manner in which they abstract the world and perform a strategic look ahead (Cross et al.,

2021, Buckner, 2018). Though this may or may not be true, it can be agreed upon that there exists

a notable list of sizable limitations that prevent AI tools from demonstrating the fluid intelligence

that makes humans the exceptional general learners they are. Though there exist numerous

issues in AI, we believe the most notable limitations are; 1) long training times (Schrittwieser et

al., 2020, Chowdhery et al., 2022, Berner et al., 2019); 2) the impotence to chart a novel path to

better guide learning; and 3) the inability to generalize to new or increasingly complex, uncertain,

and rapidly changing domains (Poggio et al., 2019, Geirhos et al., 2020). The scheme of learning

and retaining task-specific data which makes deep networks so successful, also limits it by forcing

networks to use a computationally intensive data-driven tabula rasa approach rather than allow

models to build upon what they know. Problems with this rigidity has not only lead to long training

times but also in complex domains has resulted in the daunting investigation of the trustworthiness

and reliability of these networks and the understanding they truly have of the task which they’re

undertaking (Geirhos et al. 2020, Hubinger et al., 2019, Koch et al., 2021).

On the other hand, it is believed natural intelligence, be it humans or animals, are either born with

or soon after develop innate knowledge which they can use to build increasingly complex

hierarchical representations of their environment that unfold over time. Though what this

knowledge is, and how it has been evolutionary encoded is debated, it suggests that natural

intelligence does not abide by either a blank slate or a tabula rasa approach (Wellman & Gelman,

1992, Lake et al., 2017, Velickovic et al., 2021, Silva & Gombolay, 2021). Furthermore, it is

becoming increasingly well accepted that the brain takes in the world by breaking it down into its

smallest components, even though neuroscientists are unsure how inputs are processed to create

internal models. All that is known, is that these smallest components or concepts as they are often

called (van Kesteren et al., 2012; Gilboa & Marlatte, 2017; van Kesteren & Meeter, 2020), are

hierarchically combined with increasing complexity to generate an internal model of the

environment. It has been widely hypothesized that the brain is able to do so by organizing

information into global gradients of abstraction. These gradients in accordance with a relational

memory system inclusive of processes, such as maintenance, gating, reinforcement learning,

memory, etc., continually update, store, recombine, and recall information that unfolds and is

strengthened over time to form generalized structural knowledge (Whittington et al., 2019).

A secondary backbone of intelligence is the ability to contextualize the world into an associative

framework of increasing complexity and actionable goals and subgoals (O'Reilly et al., 2014,

O’Reilly, R. C. et al., 1999, Reynolds & O’Reilly, 2009, O’Reilly, 2020). As argued in Hubinger et

al., 2019 and Koch et al., 2021, current deep networks establish objectives to assign to an

optimizer but then utilize a different model tasked with carrying out actions. This results in an

optimizer with an assigned objective, that then in turn optimizes a model that can act in the real

world. This type of architecture not only leads to problems in matching goal alignment between

creator and creation but also results in problems in trust and explainability. For example: the

authors placed an agent in an environment where it had to find and collect keys that it must use

to open a chest and gain reward. The difference between the training and testing environments

was the frequency of the objects, with the training set having more chests than keys and the

testing having more keys than chests. It was found that this simple difference in environment was

strong enough to force the agent to learn a completely different strategy than intended, i.e. finding

keys is more valuable than opening chests. This result occurred unsurprisingly as the agent

valued keys as a terminal goal and not a sub-goal. Thus, in testing, the agent not only collected

more keys than it could use but also repeatedly circled the area of inventory keys displayed on

the screen. These types of unintended strategies commonly seen by deep networks when

deployed in an environment outside of its training set have resulted in an ever-growing concern

and increasing skepticism of the outputs of deep networks.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

BridgingtheGapbetweenArtificialIntelligenceandArtificialGeneralIntelligence:ATenCommandmentFrameworkforHuman-LikeIntelligenceAnantaNair1,2andFarnoushBanaei-Kashani11.UniversityofColorado,Denver,2.DellTechnologiesIncAbstractThefieldofartificialintelligencehasseenexplosivegrowthandexponentialsuccess.T...

展开>> 收起<<

Bridging the Gap between Artificial Intelligence and Artificial General Intelligence A Ten Commandment Framework for Human -Like Intelligence.pdf

共51页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Bridging the Gap between Artificial Intelligence and Artificial General Intelligence A Ten Commandment Framework for Human -Like Intelligence

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: