5 AI Trends in 2021

52021年人工智能发展趋势

2021-02-23 17:50 TAUS

本文共1621个字,阅读需17分钟

阅读模式 切换至中文

In 2020, when the global pandemic struck the world, humanity had a groundbreaking learning experience that has forever altered almost every aspect of human interaction. But even before then, artificial intelligence (AI) and the branch of AI known as machine learning (ML) were already on the path to fundamentally reshape almost every industry. The year 2021 is ready to observe important ML and AI trends that are likely to reshape our economic, social, and industrial norms. According to a Gartner survey done in 2019, around 37% of all companies reviewed were utilizing some type of ML in their business. It is anticipated that around 80% of modern advances will be founded on AI and ML by 2022. With the surge of interest in AI technology, various new trends are ascending in the field. In this article, we will focus on five AI trends thought to be commonly used and studied in 2021 that will have a powerful impact on the language industry and beyond: vokenization, data monetization, Deep Reinforcement Learning (DRL), Automated Narrative Text Generation, and Data Science as a Service (DSaaS). 1. Vokenization Language models have been trained only with text so far. To change that, researchers from the University of North Carolina, Chapel Hill have developed a new technique called “vokenization”, which gives language models like GPT-3 the ability to “see” by enabling them to process visual input. Research into how to combine language models and computer vision has been rapidly growing as AI models that can process both linguistic and visual input can lead to highly practical applications. For instance, if we were to build a robotic assistant, it would need both computer vision to navigate the physical environment and language processing abilities to communicate about it with humans. The term “voken” was invented as part of the research. It is derived from the word “token”, which denotes a sequence of characters commonly used in language model training that roughly corresponds to a lexical unit such as a single word. And the UNC researchers call the image linked with each token in their visual-language model a “voken”. “Vokenizer” is then what they call the algorithm that matches vokens to tokens, and vokenization is the word that describes the whole process. The researchers created visual and word embeddings with MS COCO to train their vokenizer algorithm. After the training, the vokenizer managed to match vokens to the tokens in English Wikipedia. The algorithm only found vokens for about 40% of the tokens. However, that still amounts to 40% of a dataset that includes approximately 3 billion words. With this new dataset, BERT, which is the open-source, transformer-based language model developed by Google, was retrained. The researchers then tested the retrained BERT on six different language comprehension tests. The retrained version of BERT ended up performing better in all six languages. Vokenization is seen as one of the most recent breakthroughs in AI as it manages to connect text with other modalities. It’s now conceptually possible to build a robot that can not only talk but also hear and see. 2. Data Monetization With the realization that almost anything can be datafied, businesses now question how to generate value out of the high volumes of data available to or owned by them. The value of data can grow in parallel with the ways it is utilized and the insights are drawn from it. Data monetization refers to the process of generating economic value out of data. At the core of data monetization strategies that have become a growing area of interest lies treating data as an asset and maximizing its value. Being a newly explored business stream, data monetization has become many CIOs’ primary challenge. Many companies lack a business-centric value engineering methodology to identify and prioritize where and how AI/ML can derive new sources of customer, product, and operational value. With increasing successful data monetization examples, organizations begin to appreciate the economic value of data and analytics. One way to overcome this challenge is to be open to new opportunities. Marketplace models seem to be emerging as a way to provide data monetization platforms. TAUS Data Marketplace and SYSTRAN Marketplace can be named as such opportunities for language data monetization. The TAUS Data Marketplace offers the chance to any data owner to monetize their assets by making sure the seller’s data assets reach the potential buyer without any marketing effort on their part. hbspt.cta._relativeUrls=true;hbspt.cta.load(2734675, 'a25cdac6-be9f-443a-adea-8adc72fd6e87', {}); 3. Deep Reinforcement Learning Based on the recent developments in the area of Artificial Intelligence, a promising new technology for building intelligent agents has evolved according to the research done by Ruben Glatt, Felipe Leno da Silva, and Anna Helena Reali Costa at Escola Politecnica da Universidade de Sao Paulo. This new technology is Deep Reinforcement Learning (DRL) and it represents the combination of classic Reinforcement Learning (RL) which uses sequential trial and error to learn the best action to take in every situation and modern Deep Learning approaches which can evaluate complex inputs and select the best response. Currently, DRL is one of the top trending fields in Machine Learning because it can solve a wide range of complex decision-making tasks. Before the advances in DRL, solving real-world problems with humanlike intelligence had been deemed unattainable for a machine. These new approaches still require long training times for difficult and high-dimensional task domains. Generalization of gathered knowledge and transferring it to another task has been the topic of classical RL, yet it continues to be an unresolved problem in the DRL field. However, it has recently been uncovered in research that for policies, end-to-end learning is possible with high-dimensional sensory inputs in challenging task domains such as arcade game playing. Besides academic research, cases where DRL defeated humans at games, were widely covered by the media. For instance, AlphaGo defeated the best professional human player in the game called Go. AlphaStar beat professional players at the game called StarCraft II. OpenAI’s Dota-2-playing bot beat the world champions in an e-sports game. When it comes to business, many DRL applications can be seen in a wide range of sectors from inventory management, supply chain management, warehouse operations management, and demand forecasting to financial portfolio management. In Natural Language Processing, DRL is commonly used for tasks including abstractive text summarization, chatbots, and so on. A considerable amount of research is also being done on applications of DRL in education, healthcare, smart cities domains. 4. Automated Text Generation (ATG) Artificial Intelligence has been used to provide solutions to various tasks pertaining to humans in hopes to reduce cost, time and effort, among other goals, that have led to rediscovering the use of computers. One of the most intriguing questions that remains is “Can a computer be creative?”. Brian Daniel Herrera-González, Alexander Gelbukh and Hiram Calvo from the Center for Computing Research analyze this question in their relevant study. One of the tasks that AI was applied to is the automatic generation of stories. Although certain patterns have been defined, there is still no clear methodology in place. So far, based on academic research, two approaches regarding the task of automatic text generation have been identified: the symbolic approach and the connectionist approach. In the symbolic approach, one of the advantages is that it’s easy to delimit the path followed by the actions of the story whereas the main disadvantage is that it lacks creativity and creates stories that are repetitive with a reduced repertoire of events. In the connectionist approach, the problem of lack of novelty in stories is resolved, however, the coherence is lost in the story. In the field of automated text generation, solutions such as sequence-by-sequence models (seq2seq) with recurrent neural networks (RNN) and unsupervised learning by generative adversarial networks GAN or Bidirectional Associative Memories have been proposed and discussed. When it comes to the practical use of ATG, it’d be fair to say that most contemporary internet users come across ATG output daily. Among the established applications for automated text generation are football reports, earthquake notifications, stock exchange news, weather, and traffic jam reports. As the field advances in its application, it gets harder to differentiate between automatically generated articles and those written by humans. According to tests with football match reports, readers rated the output of the ‘robot journalist’ to be more natural than those of a ‘real’ editor. One interesting example of ATG is the new Brothers Grimm fairy tale produced by AI where the algorithm was trained to mimic the literary style of the Brothers Grimm. Additionally, ATG is also widely used in e-commerce where large quantities of product descriptions are produced by ATG with greater ease. 5. Data Science as a Service (DSaaS) Data Science as a Service (DSaaS) involves the delivery of data gathered through advanced analytics applications to corporate clients, run by data scientists at an outside company (service provider). The aim is to collect business intelligence and make informed decisions about new business efforts. The way it works is: clients upload the data to a big data platform or cloud database. The data scientist and service provider team specifically incorporated with data engineers work on the uploaded data. They analyze the data and prepare a report on which company is likely to buy your products, your rival details, your net earnings, revenue, etc. Data Science as a Service is mostly used by companies that are suffering due to the shortage of data scientists. Eventually, DSaaS helps in providing growth in the business. hbspt.cta._relativeUrls=true;hbspt.cta.load(2734675, 'c6f061a3-1d57-4f4f-a1d3-42b833a8b1f9', {});
2020年,当全球大流行病袭击世界时,人类开创了一种学习经验,这种经验会永远地改变人类互动的几乎每一个方面。但即便在那之前,人工智能(AI)和被称为机器学习(ML)的AI分支已经走上了从根本上重塑几乎每一个行业的道路。 2021年是观察重要的ML和AI趋势的一年,这些趋势很可能重塑我们的经济,社会和工业规范。根据Gartner在2019年所做的一项调查,约37%的受访公司在其业务中使用了某种类型的ML。预计到2022年,大约80%的现代化先进设备将建立在人工智能和ML之上。 随着人们对人工智能技术兴趣的激增,该领域出现了各种新的趋势。在本文中,我们将重点关注被认为在2021年受到广泛使用和研究的五个人工智能趋势,它们将对语言行业及其他领域产生强大影响:语音化、数据货币化、深度强化学习(DRL)、自动叙事文本生成和数据科学即服务(DSaaS)。 语音化 到目前为止,语言模型只通过文本进行训练。为了改变这种情况,来自教堂山的北卡罗来纳大学的研究人员开发了一种名为“vokenization”的新技术,通过使GPT-3这样的语言模型能够处理视觉输入,从而赋予它们“看”的能力。随着能够同时处理语言和视觉输入的人工智能模型的高度实用应用,如何将语言模型和计算机视觉结合起来的研究正在迅速发展。例如,如果我们要建造一个机器人助手,它既需要计算机视觉来导航物理环境,也需要语言处理能力来与人类进行沟通..。 " voken "这个词是研究的一部分。它来源于单词“token”,token表示语言模型训练中常用的一组字符,大致对应于一个词汇单位,如单个单词。北卡罗来纳大学的研究人员将他们视觉语言模型中与每个符号相连的图像称为“voken”。“Vokenizer”就是他们所说的将vokens与token匹配的算法,vokenization就是描述整个过程的单词。 研究人员用COCO女士创建了视觉和单词嵌入,以训练他们的vokenizer算法。训练结束后,赋音者成功地将英语维基百科中的元音与符号匹配起来。该算法仅为40%的令牌找到了vokens。然而,这仍然相当于一个包含约30亿个单词的数据集的40% 有了这个新的数据集,BERT(谷歌开发的基于变压器的开源语言模型)得到了再训练。然后,研究人员对伯特进行了六种不同的语言理解测试。经过再训练的BERT在所有六种语言上的表现都更好。。 语音化被视为人工智能最新的突破之一,因为它成功地将文本与其他形式联系起来。现在,从概念上讲,制造一个不仅会说话,还会听会看的机器人是可能的 2.数据货币化 由于认识到几乎任何东西都可以被数据化,企业现在开始质疑如何从他们可用或拥有的大量数据中产生价值。 数据的价值可以随着数据的利用方式以及从数据中得出的见解而同步增长。数据货币化是指从数据中产生经济价值的过程。数据货币化策略的核心是将数据视为一种资产并使其价值最大化,这已成为一个日益增长的兴趣领域 数据货币化作为一种新兴的业务流,已成为许多首席信息官面临的主要挑战。许多公司缺乏一种以业务为中心的价值工程方法,来识别和确定AI/ML在何处以及如何获得新的客户、产品和运营价值来源。随着成功的数据货币化例子的增加,企业开始重视数据和分析的经济价值 克服这一挑战的一个方法是接受新的机会。市场模式似乎是提供数据货币化平台的一种方式。TAUS数据市场和SYSTRAN市场可以被称为语言数据货币化的机会。TAUS数据市场为任何数据所有者提供了将其资产货币化的机会,方法是确保卖方的数据资产在不进行任何营销努力的情况下接触到潜在买家。 hbspt.cta._relativeURLS=true;hbspt.cta.load(2734675,'A25CDAC6-BE9F-443A-ADEA-8ADC72FD6E87',{}); 3.深度强化学习 基于人工智能领域的最新发展,根据圣保罗理工大学的Ruben Glatt、Felipe Leno da Silva和Anna Helena Reali Costa所做的研究,一种很有前景的构建智能代理的新技术已经发展起来。这种新技术就是深度强化学习(DRL),它代表了经典的强化学习(RL)和现代深度学习方法的结合,后者使用连续的尝试和错误来学习在任何情况下采取的最佳行动,而现代深度学习方法可以评估复杂的输入并选择最佳响应。目前,DRL是机器学习领域最热门的领域之一,因为它可以解决广泛的复杂决策任务。在DRL技术取得进展之前,用类似人类的智能来解决现实世界中的问题被认为是机器无法实现的。 这些新的方法对于困难和高维的任务域仍然需要很长的训练时间。对所收集到的知识进行泛化并将其转移到另一个任务中一直是经典RL的主题,但这一直是DRL领域中尚未解决的问题。然而,最近的研究发现,对于政策而言,端到端学习是可能的,在具有挑战性的任务领域,如街机游戏玩法,通过高维度的感觉输入。 除了学术研究,DRL在游戏中击败人类的案例也被媒体广泛报道。例如,AlphaGo在一种叫做围棋的游戏中击败了最好的职业人类棋手。AlphaStar在名为《星际争霸2》的游戏中击败了职业选手。OpenAI的Dota-2机器人在一场电子竞技游戏中击败了世界冠军。 当涉及到业务时,从库存管理,供应链管理,仓库运营管理,需求预测到金融投资组合管理等广泛的部门中都可以看到许多DRL应用。在自然语言处理中,DRL通常用于抽象文本摘要,聊天机器人等任务。DRL在教育,医疗,智慧城市等领域的应用也得到了大量的研究。 4.自动文本生成(ATG) 人工智能已被用于为各种与人类有关的任务提供解决方案,以期减少成本、时间和精力,以及其他目标,这些目标导致重新发现计算机的用途。现在最有趣的问题之一是“电脑能有创造力吗?”来自计算研究中心的Brian Daniel Herrera-González, Alexander Gelbukh和Hiram Calvo在他们的相关研究中分析了这个问题。人工智能应用的任务之一是自动生成故事。虽然已经定义了某些模式,但仍然没有明确的方法。 迄今为止,基于学术研究,关于文本自动生成的研究主要有两种方法:符号方法和连接方法。 在象征性方法中,它的一个优点是很容易界定故事行动所遵循的路径,而它的主要缺点是缺乏创造性,创造的故事是重复的,事件数量较少。在连接主义的方法中,故事缺乏新颖性的问题得到了解决,然而,故事失去了连贯性。在自动文本生成领域,已经提出并讨论了一些解决方案,如使用递归神经网络(RNN)的序列序列模型(seq2seq)和使用生成对抗网络GAN或双向联想记忆的无监督学习。 说到ATG的实际应用,公平地说,大多数当代互联网用户每天都会遇到ATG输出。自动生成文本的应用程序包括足球报告、地震通知、股票交易新闻、天气和交通堵塞报告。随着该领域应用的发展,人们越来越难以区分自动生成的文章和人工撰写的文章。根据对足球比赛报道的测试,读者认为“机器人记者”的报道比“真实的”编辑的报道更自然。ATG的一个有趣例子是人工智能制作的新《格林兄弟》(Brothers Grimm)童话,算法被训练成模仿《格林兄弟》的文学风格。此外,ATG也广泛应用于电子商务,在电子商务中,大量的产品描述可以由ATG更容易地生成。 5.数据科学即服务(DSaaS) 数据科学即服务(DSaaS)涉及将通过高级分析应用程序收集的数据交付给企业客户,由外部公司(服务提供商)的数据科学家运行。其目的是收集商业情报,并就新的商业努力做出明智的决策。 其工作方式是:客户端将数据上传到大数据平台或云数据库。数据科学家和服务提供商团队专门与数据工程师合作处理上传的数据。他们对数据进行分析,并准备一份关于哪家公司有可能购买你的产品,你的对手细节,你的净收益,营收等的报告。 数据科学即服务多被那些因数据科学家短缺而苦不堪言的公司所使用。最终,DSaaS有助于提供业务增长。 hbspt.cta._relativeURLS=true;hbspt.cta.load(2734675,'C6F061A3-1D57-4F4F-A1D3-42B833A8B1F9',{});

以上中文文本为机器翻译,存在不同程度偏差和错误,请理解并参考英文原文阅读。

阅读原文