ERROAR!2 2018

chatbot (interactive web-based application) 聊天机器人(互动在线应用)

The ERROAR! series speaks to the errant and noise that emerge from experiments of artificial intelligence for understanding human’s cognitive functions, and how these technological errors open up new creative potentials that in return reconfigure our perceptions, affect and imaginations. Most visual-textual-sonic materials involved are generated from the process in which I employ open data to train prevalent machine learning algorithms including RNN (recurrent neural network) and CNN (convolutional neural network).

In 2014, in order to defeat high-profile Go player Lee Sedol, AlphaGo is allegedly powered by 1920 CPUs and 280 GPUs. The first work ERROAR!#1 is a pixel art image that is made up of 1920 images of CPUs and 280 images of GPUs to materialize AlphaGo Lee’s hardware use, or its “brain capacity” on a 1-1 scale. It shows one of the key problems of current AI technologies that AI has to consume enormous electricity and computing power in order to learn from large datasets (memories and experiences).

ERROAR!#2 visualizes machine-imagined news headlines that are trained from Toutiao’s open database. It reveals the incomplete states of machine learnt knowledge due to lack of training time and computing resources. In the style of Toutiao headlines, these generated language is almost incomprehensible for humans, without following the common use of Chinese syntax. Inventing new cultural myths, the illegibility of the machine-imagined headlines leaves silent roar in minds of the readers.

ERROAR!#3 is a three-channel video essay that traces the economy of artificial intelligence and how its applications have co-mingled with the development of human living conditions. The video revives a desktop scene — clicking through hyperlinked web pages to gradually trigger open over 50 browser windows embedded with video clips and generative software, which are all engineered by the artist. From the birth of one single pixel on a screen, to the algorithm-generated visuals enabled by inceptionism, the video falls into maelstroms and rainbow spectrum, an ouroboric deep dream. This work aggregates a large amount of visual spectacles and fantasies of AI and technologies in popular culture and stock photograph database. The sound work accompanied to the video is also composed with machine generated sonic materials to enhance the pareidolia, representing a hyper-reality of our collective fear and resistance, hope and reflections.

ERROAR!#4 contemplates on the phenomenon of “cannibalism” as a metaphor for cultural and technological assimilation, especially in the process of human-artificial-intelligence coupling. Entirely composed of stock videos, the single-channel video essay is derived from an online anecdote about the first case of “virtual cannibalism” conducted by AI agents during DARPA’s early experiments. The piece of generative poetry titled “Past Said Lore” on the paper scroll is co-authored with a recurrent neural network to write in the style of John Milton’s Paradise Lost. Through a webcam and a mirror, the networked software “misrecognizes” human faces as “objects”, due to incorrect and incomplete training and learning. Being objectified by an object in error creates an infinite feedback loop between the mirror and the camera, performing a metaphorical autophagy, just like the closed loop of human skeleton eating tailbone.

Related text: Interview with IOAM in Chinese 与中间美术馆的访谈

人们通过构建人工神经网络来开发智慧生命的同时进一步理解了人类自身的认知形成。ERROAR! 观察放大这些实验中产生的误差和噪音,以及这些技术错误如何打开了新的创造潜能,从而逆向配置我们的感知、情动(Affect)和想象力。我使用不同格式的开放数据来训练盛行的机器学习算法如RNN(递归神经网络)和CNN(卷积神经网络),所生成的结果交织出了混合媒体的装置。ERROAR!强调算法系统的内部结构。这些系统动员了当代图像、声音和文本的生产,质问我们一直联体共生且相互作用的网络化的赛博格现实。

2014年,为了击败职业围棋手李在石,AlphaGo据称需要1920个中央处理器(CPU)和280个图像处理器(GPU)。ERROAR!第一号 是由1920张CPU图像和280张GPU图像组成的像素画,以1:1比例还原了AlphaGo的硬件使用大小,即该智能主体的“脑容量”。借此指出了当前人工智能技术的关键问题之一 — 人工智能必须消耗巨大的电力和计算能力才能从大型数据集(记忆和经验)中学习。

ERROAR!第二号是一个互动网页装置(聊天机器人)。机器人的聊天语料是从《今日头条》的开放数据库训练出来的新闻标题,即由机器想象,不存在的语言。虽然看似头条标题的文字风格,但因为没有遵循汉语的普遍用法,这些算法生成的标题几乎是无法理解的,揭示了由于缺乏训练时间和计算资源而导致机器“习得知识”的未完成状态。机器想象的语言的不可读性激发了新的文化迷思,在读者的意识中留下了无声的巨响。

ERROAR!#3 是一部三通道的录像,追踪人工智能的经济,以及智能应用如何与人类的生活状态耦合发展。录像重现了一个电脑桌面的场景:从一个网页程序陆续点击打开超过50个浏览器窗口。这些超链接的网页中嵌入了由艺术家开发制作的录像和动画片段以及程序应用。从屏幕上单个像素的诞生,到计算机依据开始主义(inceptionism)生成的算法作画,录像逐渐坠入旋涡和彩虹频谱,一个乌洛波洛斯式的深梦(deep dream)。录像中所用的素材汇集了大量大众文化及库存摄影图片库中人们对人工智能和科技的幻想奇观。伴随录像的声音作品用类似的方法合成了机器学习的声音素材,更加增强深梦的空想性错觉,再现了集体恐惧和抵制、希望和反思的超真实景观。

ERROAR!#4 将“食人”现象看作是文化和技术同化的隐喻,特别是在人-人工智能耦合的过程中。三个部分 – 单频录像完全由网络收集的图库视频组成。脚本基于一件关于人工智能训练错误的网络轶事,讲述DARPA(美国国防高级研究计划局)早期研究智能主体的实验。这也是第一个“虚拟食人”的案例;《过去说过的话》的卷轴是基于弥尔顿《失乐园》与一个递归神经网络合著的生成式诗歌,自动无休地写作与印刷;通过网络摄像头,联网的软件因不充分的学习和训练而只能将人脸识别为物件。“被物错误地物化”在镜面与网络摄像头之间反馈,像是自己在啃食自己,细胞自我吞噬。