Negative optics in vision machines

Luciana Parisi/Texts/Essays/Negative optics in vision machines.pdf

P. 1
AI & SOCIETY https://doi.org/10.1007/s00146-020-01096-7 ORIGINAL ARTICLE Negative optics in vision machines Luciana Parisi1 Received: 23 July 2020 / Accepted: 14 October 2020 © Springer-Verlag London Ltd., part of Springer Nature 2020 Abstract Can machine vision step beyond the ocularcentric metaphysics of the Western gaze and the reproduction of racial capital? Paul Virilio argued that machine vision requires no perceptual response or recognition of the world. The computer’s series of coded impulses mediate the real autonomously from physical and energetic analogies. This article calls “negative optics” this inhuman mode of machine vision that withdraws from the ocularcentric rules of transparency. Not only negative optics offers an internal critique of ocular metaphysics, but it also defies the equation of value between 0 and 1 s that sustains the universal law of capital. The equation of value needs the nullification of the racialized and gendered body to meet the function of surrogate machines feeding the neo-liberal human subject. Here, however, the negative optics of machine vision withdraws matter from the equation insofar as the AI mismatches of concepts and objects requires the formation of sociotechnical assemblages amongst surrogates of all kinds. The intricacies of this surrogacy are discussed in connection to artist Elisa Giardina’s video installation entitled, The Cleaning of Emotional Data (2019). This video presents the background to address a surrogate human–machine alliance that steps beyond the human–machine equation of value. By starting from the negativity of the image, the racialized and gendered conditions of techno-social labor under artificial intelligence capitalism show that the equation of value maintains the condition of the zero of blackness, which like matter without form, has no value (On matter beyond the equation of value 2017). Drawing on François Laruelle, the article continues to elaborate the possibilities of a material image without form in terms of what Laruelle calls, the “fractal algorithm of the photo”. The article concludes that the negative optics of machine vision stands for the alien origination of knowledge, values, materialities that overturn the equation of value through the fractal infinities of 0 s. Keywords Negative optics · The equation of value · Generative adversarial networks · Surrogacy · Auto-imagining · Matter without form · Blackness · Non-photography Already in the late 1980s, Paul Virilio argued that the technical image in computational and cybernetic machines could no longer be understood according to the framework of ocularcentric metaphysics and the epistemological rules of the relation between vision, knowledge and power. If Virilio suggested that the vision machine no longer coincides with how humans see, it is because he already foresaw the overturning of the Western gaze of the Platonic model of shedding light on the darkness of matter. With the automation of knowledge, the world view of technological universalism extends the Western gaze beyond what can be enlightened. In a computer, the optically active electrons of machines correspond to a series of coded impulses that mediate the real * Luciana Parisi l.parisi@duke.edu 1 beyond physical or energetic analogy. As Virilio claimed, with the “automation of perception” (1994) image feedback is no longer assured by the interaction with the world, insofar machine vision does not shed light on dark matter. As a non-dialectical medium, machine vision requires no relation with or response from the world to exist and to function as a data processor. In other words, with the automation of knowledge, we have a programmed perception that is no longer based on observation and reflection of the object observed. With machine vision, it becomes evident that the feedback function of algorithms incorporates the world in terms of input data through which the world is predicted and acted upon in anticipation of its happenings. Following Virilio, one can suggest that a negative optics as opposed to an enlightened visibility comes to redefine vision in terms of a mediatic function that does not rely on light. Duke University, Durham, NC, USA 13 Vol.:(0123456789)
P. 2
AI & SOCIETY Virilio understands this negative opticality in terms of the computer being blind or un-affected by light. “Blindness is thus very much at the heart of the coming ‘vision machine’” (1994, p 72). The latter coincides not with a picture frame, but with a statistical calculation and a pixilation of the world in terms of binary language, or spatial discretization for which sets of coded impulses become increasingly intensive and correspond to infinitely small durations (1994, p 72). According to Virilio, the computational function of media breaks away from ocularcentric metaphysics because the “absolute-speed machine” (1994, p 72) breaks from the extensive geometric optics defined by observables and non-observables, enlightened and dark objects. It is this speed that defines the intensive time of a “vision machine” that does not look at the world but instead generates an avalanche of inputs at the same point, and between points as if it were produced by a “short concentration spans by means of surprise” (1994, p 72). In other words, inputs are statistically generated along with a series of short spans, intensive durations that are as it were disentangled from each other and yet they appear united by the speed of algorithmic functions that allows machines not to see. However, what Virilio calls blindness of machine vision is not to be understood in terms of machines seeing the world in terms of a transparent or neutral I that programs the world. Instead, one can argue that blindness here coincides with a techno-political theorization of machine vision that starts from machines’ negative optics as involving an anti-ocularcentric practice that breaks from the self-positing of the metaphysics of representation. If the gaze and its mediatic extensions remain a tool for un-veiling the other to impart, extend, consolidate the selfdetermination of the transcendental eye/I, negative optics instead stays with the unsubstantial, unformed, unvalued dimensions of matter, a practice of subtracting light from the surface of the image in which the self-determining gaze continues to mirror himself. As this article will attempt to explain, what Virilio calls the blindness of the “vision machine” has demarcated a series of internal ruptures in the optical regime of representation and, this article argues, in the universal model technology. On the one hand, these fragmentations have been understood in terms of an extension of the universal model of technology whereby statistical prediction aims to reduce the world to one truth thus replacing the ocular model of truth with statistical functions. On the other, the negative optics of computation processing can be taken to work as an instance of how the inhuman mode of vision challenge the dialectic of visibility and transparency and thus offers an internal critique of ocular metaphysics and its entanglement with colonial and racial capitalism (Lowe 2015).1 1 According to Lisa Lowe, the concept of “racial capitalism” developed by Cedric Robinson, implies that capitalism expands not through rendering all labor, resources, and markets across the world 13 This article suggests that the architecture of racial capitalism also contains within its articulations a fundamental split between the human and the machine based on the equation of value that functions to re-introduce the transcendental matching of concepts and objects on machinic perception. In other words, the ocularcentric equation of value at the core of racial capitalism imparts a universal model of technology at once merging and opposing humans and machines, whereby racialized humans are used to train machines to correct their negative optics. In other words, not only ocularcentrism becomes technologically extended in the performative operation of mediatic surveillance, but racialized capital also grants that this is reproduced by racialized humans training blind machines. This automated practice of extraction of value requires that humans train machines to correct their learning patterns and follow the optical order that matches transcendental concepts with objects. It has been argued that these highly underpaid jobs contracted through online platforms such as Amazon Mechanical Turk2 are a new form of racialized surrogacy whereby humans are asked to teach machines how to read images by following the order of pan-opticality of the self-determining subject. According to Neda Atanasoski and Kalindi Vora, the technoliberal articulation of today’s racial capitalism works to conceal the uneven racial and gendered relations of power articulated with and through machines. At the core of technological innovation therefore lies a surrogate relation of technology towards the “human sphere of life, labour and society” that enables, in turn, the constant re-constitution of the liberal subject (Atanasosky and Vora 2019, p 10). What they call “the racial unfreedom of the surrogate” can contribute to explain how the ocularcentric order by mirroring into the world, finds its own self-determining humanity, or as the Footnote 1 (continued) identical, but by precisely seizing upon colonial divisions, identifying particular regions for production and others for neglect, certain populations for exploitation and still others for disposal (Lowe 2015, p 149). 2 Amazon’s Mechanical Turk, as well as CrowdFlower, Clickworker, Toluna, and others, are largely unregulated websites that allow businesses and individuals to post short tasks that are also called Human Intelligent Tasks assignments and pay workers—in cash or, sometimes, gift cards—to complete them. In particular, Amazon set up the website Mechanical Turk in 2005 for humans to perform tasks that are hard for computers. Amazon executive Jeff Bezos has called this kind of human task, “artificial artificial intelligence.” This means that when a task is easier for a human than a computer, the computer calls a human. It is this double artificiality with which underpaid humans are identified that will be interesting to explore in this new condition of extraction of value. Amongst the most common tasks are the recording of information appearing in an image and the transcription of audio and video files. See https​://www.pewre​searc​h.org/inter​ net/2016/07/11/what-is-mecha​nical​-turk/ (last accessed June 20th, 2020).
P. 3
AI & SOCIETY authors say, by drawing on Hortense Spillers, a project for “feeling human” that sustains the epistemological ground of racial engineering.3 It is my argument that as much as the negative optics of machine vision entails the filtering out of adversarial patterns, namely the machinic mismatch between concepts and objects, the equation of value also shows how surrogate humans are entangled with surrogate machines in the formation of socio-technical assemblages under neo-liberal racial capitalism. An instance of the automated infrastructure of racial capital that exposes the equation of value between surrogate humans and machines can be found in artist Elisa Giardina’s video installation entitled, "The Cleaning of Emotional Data" (2019). As it may become clearer later, the cleaning of data coincides with a capital re-introduction of the ocular metaphysics within computational processing where Amazon Mechanical Turk (MTurk) have the task of automating perception according to the model of knowledge founded on verification (or proof of recognition) and the equation of value of human–machine. What Amazon Mechanical Turk allows is to turn the speed or the intensive time of computation into the extensive time of optical perception. From this standpoint, one could ask, how can the blind vision of computation processing offer an internal critique of the colonial capital and its ocular metaphysics? How can the human–machine alliance go beyond the human–machine equation of value starting from where working with Amazon Mechanical Turk can become a locus of theorization of a machine epistemology, which takes cultural, affective, aesthetic labor to counter-actuate the universality of technology—and its colonial dialectic of the visible and the invisible, the optical and non-optical, the one and zero? To what extent can the negativity of blindness contribute to expose the internal crisis of ocularcentric capital, and what forms of epistemological capacities can rather originate from the negativity of the image entangled with the racialized conditions of techno-social labor? This article will point out that debates about the aesthetic possibilities of machine vision based on the opposition between optical human perception and non-optical automated perception seem to reinforce, instead of radically challenging, the model of metaphysical decision correlating 3 The authors draw on black feminist, Hortense Spillers, to discuss the surrogate human effect at a constitutive part of the grammar of colonialism and technoliberalism. I take these critical reflections about how the racialized other as much as the machine have surrogate status in colonial technocapitalism as central to the arguments about the racialized socio-technical assemblage at play within contemporary images of AI and automation. See Hortense Spillers, “Mama’s Baby, Papa’s Maybe: An American Grammar Book.” Diacritics 17 (1987),p 67. knowledge and ocularcentrism at the core of Western philosophy. The aesthetico-political critique of machine vision today has been focusing on the shift from the representational to the operational image, defining the technical image in terms of a programmed action—a cybernetic model of steering conduct. As discussed in this article, the proliferation of operational images also corresponds to the formation of an aesthetic reconfiguration of control that explains the correlation of vision and knowledge directly in terms of behavioral conduct. The operational image, therefore, together with Virilio’s notion of the vision machine or blind vision point to the cybernetic nature of the mediatic, technical image as the steering of conducts towards pre-programming responses. Similarly, the non-ocular functions of machine vision resonate with the arguments that machines produce invisible images that are only readable by machines, but not by humans. This article will particularly dwell with this double level of argumentation, which, on the one hand, tells us that the negative optics of machine vision mainly leaves humans out of the loop—disconnected from the very matrix of communication upon which we rely, and on the other, points to how the blindness of machines rather relies on a racialized equation of value at the core of techno-capitalism, for which blackness like matter (Ferreira da Silva 2017) has no value and constitutes the surrogate humanity merged with the zero value of machines, whose invisible images are set to be re-programmed by the ocularcentric knowledge. From this standpoint, this article argues that one other possible elaboration of negative optics that can challenge the human–machine equation of value (while re-affirming the universal agency of the Man) and rethink the alliance between the racialized zero value of surrogate humans and machines, is to borrow from François Laruelle’s “non-philosophy” (2010) which concerns the question of dark optics and immanent vision. In particular, Laruelle’s claim about the fractal algorithm of the photo will be explored to discuss the configuration of a negative optics in automated vision in recent efforts to clean negative randomness in Generative Adversarial Networks. The article suggests that the negative materiality of the computational image does not only show that knowledge can be divorced from ocularcentrism, but also that the invisible image of machines is part of alien epistemologies that overturn the equation of value with the infinities of 0 s. This article proposes that the material processing of randomness in computation is part of the expansion of heretic epistemologies that start from dark optics and address what Denise Ferreira Da Silva calls blackness, namely matter without form, or “matter beyond the equation of value” (2017). The materiality of the computational compressing of infinities, however, coincides not with the physical structure of the machine, but with its abstract mode of operation—the 13
P. 4
AI & SOCIETY algorithmic learning to learn from the negativity of randomness—entailing both the elaboration of indeterminacy and the incompleteness of systems. Matter without form coincides not with invisible images without meaning, but with a real that can only be cloned as auto-impressions of dark optics, stemming from within the infinite discretization of opacities, the fractal singularizations of blackness in artificial visions. From this standpoint, this article suggests that the aesthetico-political engagement with machine vision must be divorced from ocular epistemologies and the critique of vision. To argue that matter without form entails the negativity of algebra is to propose that matter like infinities break with the equation of value, and thus refuses the metaphysical equivalence of knowledge and vision that sustains the surrogate subjection of humans and machines. The blind vision of computational processing rather speaks of the challenge of reinventing epistemology from the standpoint of dark optics of the human–machine condition of surrogacy in the operational extension of ocular capital. 1 Machine don’t see and we know it “The production of sightless vision is itself merely the reproduction of an intense blindness that will become the latest and last form of industrialisation: the industrialisation of the non-gaze.” (Virilio 1994, p 73). If one follows Virilio’s reflections on the blindness of machine vision, it is possible to suggest that this non-gaze is the result of the formation of new epistemological orders moving from the extensive time of analogical machines to the intensive time of the computer. According to Virilio, this new order incorporates the transformation of the formal logic of traditional representation into the dialectic logic of cinema and photography into the paradoxical logic of the digital image. In particular, for Virilio, as much as the extensive time of analog media enabled a continuity of experience between the past and the present, the paradoxical logic that accompanies the computer rather brings a remote object in the here and now. Here perception turns into action at a distance, a vision running at the speed of light. Interestingly, Virilio insists that the paradoxical logic of the computer allows for the future never to occur but rather disappear in the statistical programming of acting at a distance. This remote action is also central to the classical cybernetic account of remote control on the one hand, and the prediction of the future, on the other. As Virilio states: “[w]hen a missile threatening in ‘real time’ is picked up on a radar or video, the present as mediatised by the display console already contains the future of the missile’s impending arrival at its target” (1994, p 66). 13 Blindness, therefore, results from the speed of light whose increasingly intensive intervals bring forward an alien perception of the object as if the latter was coming from the future. Instead of the reciprocal presupposition between light and dark—or the dialectical presupposition of lightness vs darkness that characterized the passive optics of extensive media, such as photography—for Virilio, computer vision deploys an intensification of light, namely depending on the velocity of photons and their constant pixelation. As much as photons are abstracted in the statistical image, the rapid calculation of pixels dissimulates the future in the ultra-short time of an on-line “compunication” (computer communication) (1994, p 68). In this new “high-tech mix,” a paradoxical logic sees the fusion of the object with its equivalent image. As Virilio reminds us, these processes of real-time deception will win out over the weapons systems of classic deterrence (1994, p 68). Blindness, therefore, defines the speed of the synthetic image, where speed prevails over time and space, matter and form. Here the gaze of reason becomes a statistical thought or a statistical optics, which according to Virilio, “generates a series of ‘visual illusions’, or ‘rational illusions’, which affect our understanding as well as reasoning” (1994, p 75). Virilio warns us against the coming statistical intoxication of the technical image, claiming that society is “sinking into the darkness of a voluntary blindness” (1994, p 76) that will forever occlude the horizons of self-determining knowledge. This statistical image does not only rely on the speed of numerical calculation, but also on the speed of cognitive perception. It shows us that the medium is not a tool but more importantly a machine language, where the speed of algebraic relations has taken over the linearity of input–output communication. As the speed of computational processing has shifted from digital pixilation to the neural network architecture of machine learning, the blindness of the technical image is now being understood as a black box that carries out functions that are impossible to observe. According to artist Trevor Paglen, this new quality of the technical image can be seen in Harun Farocki’s use of computer visions in Eye/ Machine III, where, as Paglen notices, he was “trying to learn how to see like a machine” (2014). In particular, these machines seem not to be representing things in the world and thus neglect the ocularcentric correlation between eyes and objects. Instead of observing the world, computer vision, Paglen claims, does things in the world. However, one could ask, what does it mean that computer images cannot be seen and yet they do things in the world, what does this “doing” amount to? On the one hand, one could suggest that computer images are operational because they point to a certain activity of the automated image itself in as much as it communicates with other images before communicating to us. This is how machine to machine
P. 5
AI & SOCIETY communication constitute the global infrastructure of surveillance and governance. On the other hand, in addition to the automated function of communication, the operational image is also said to activate the cognitive-perceptual apparatus of the real, which precisely pre-program conducts, or predicts behavior. In other words, the operational image fits with the cybernetic imperative of self-regulatory feedback which ensures that machine-to-machine communication can condition all orders of reality. In a more recent reflection about operationality of the technical image, Paglen insists that machine to machine communication has made sure that images are now invisible. In the short article “Invisible Images (Your Pictures are looking at You)” (2016), Paglen suggests that this invisibility coincides with how the proliferation of automated vision, as for instance in the case of reverse image search engines. According to Paglen, “[i]mages have begun to intervene in everyday life, their functions changing from representation and mediation, to activations, operations, and enforcement. Invisible images are actively watching us, poking and prodding, guiding our movements, inflicting pain and inducing pleasure. But all of this is hard to see” (2016). Invisible images are at the core of new forms of identification systems, from Automatic Licence Plate Readers (ALPR), Optical Character Recognition (OCR), to Deep Face Algorithm, Facebook’s DeepMask and Google’s TensorFlow whose algorithmic patterns recognize “people, places, objects, locations, emotions, gestures, faces, genders, economic statuses, relationships, and much more” (Paglan 2016). In particular, “deep learning” networks are built out of dozens or even hundreds of internal software layers that exchange information. This is at the core of recursive feedback, where the neural network layers of the software pick apart a given image into component shapes, gradients, luminosities, and corners. Those individual components are convolved into synthetic shapes, which are compared with the images fed into the CNN (convolutional neural network), and which the network has been trained to recognize. The invisible image is then activated by software “neurons” as the network finds common patterns with other images. For Paglen, this automated synthesis of images means that power has itself become invisible and yet institutes an intensively crafted mode of control, policing and market. For instance, he conjectures that the new relation between power and invisibility implies how the specific use of metadata signature of every single person, based on race, class, the places they live, the products they consume, their habits, interests, “likes,” friends, leads to a reification of those categories at another level. Filtering out mis-matches of individualized metadata profiles become part of an automated function, whose task is to collect municipal fees, adjust insurance rates, conduct targeted advertising, prioritize police surveillance, and so on (Paglan 2016). From this standpoint, metadata signature appears to support diversity, but only because these differentiations can become subsumed to marketing and predictive policing. The invisible image, therefore, exposes the ineffective practice of the critique of representation insofar as, according to Paglen, machines are not concerned with how humans see. While there is no possibility to either subverting the invisible regime of pattern recognition or fix it with more representations, for Paglen the invisible image demarcates the beginning of a stealthy mode of power that we are yet to address (2016). However, one may argue that Farocki’s reflections on computer vision invite a reflection on how to train the human to see like a machine, and that a fundamental trick of ocularcentrism is precisely to oppose the visible to the invisible, to re-inject the truth of representation back into the deep learning of automated networks. To put it in another way, it is evident that if machines don’t comply with the oculacentric correlation between vision and knowledge, left to their own devices of searching for images by images, they will be unable to filter out dirty data and or match patterns of an image that does not exist with a given set of objects/ concepts. In CNNs for instance, an image search that does not match with the image found implies an increase in randomness or noise or what Hito Steyerl calls, “dirty data” (2015) that make the signal fuzzy and must be filtered out from communication. In particular, Steyerl discusses what Google calls “inceptionism,” by referring to the automated creation of patterns from noise in “deep dreaming,” where image recognition algorithms are not distinguished from but rather looped on randomness or noise. Steyerl argues that here signal and noise are already predetermined by pre-existing categories and probabilities that reproduce aesthetic and social relations through the immediate correlation of data and hypothetical inference. In particular, automation comes to coincide with a regime of artificial interpellation that asks users to recognize themselves by constantly registering eye movement, behaviors and preferences and thus by re-imparting the information-encoding organizational principle of segregation and discrimination on users. As the influx of data comes into the system of prediction it matches correlations across patterns and drawing branches of decision-making between unrelated information, signals and noise: namely deriving meaning from “apophenia.”4 From social scores to credit scores, from academic scores to threat scores, as well as commercial and military patternof-life observations, Steyerl sees apophenia as an upgraded 4 “Apophenia” (/æpoʊˈfiːniə/) is the tendency to mistakenly perceive connections and meaning between unrelated things. The term (German: Apophänie) was coined by psychiatrist Klaus Conrad in his 1958 publication on the beginning stages of schizophrenia. 13
P. 6
AI & SOCIETY version of Walter Benjamin’s “optical unconscious,” which reformats social hierarchies by ranking, classifying and filtering noise following a matrix of radicalized discriminations. The invisible image, therefore, manifests itself in the form of a brutal form of decisionism, which, according to Steyerl defines “a practice of data divination,” namely the search for given determinations as if emerging from an unknown divine order of knowledge (2015). However, one could ask, isn’t this divine decisionism simply an extension of the colonial epistemology which relies on the surrogacy of human and machine as fundamental to racial capital? One could argue that capital practices of the equation of value entail not simply the prediction of value through the meta-programming of data or in other words, an automation of value derived from functions that successfully carry out tasks. Instead, divine decisionism relies less on given functions that conform to the ocularcentric dialectic, and more on the split between surrogate humans and machines in the praxis of the equation of value, where the filtering out of noise foregrounds the practice of extraction. The optical matching between blind images, transcendental concepts and empirical objects is not given but is rather constantly re-introduced in computer vision by capital’s exceptional decisionism. The latter can only dissimulate the ingression of the real in the vision machine, and rather continues to make efforts to pre-program results to preclude the computational process of compression to run with the negativity of the function—the increasing noise or randomness that cannot be automatically translated into signal or pattern. But how does this strategy of dissimulation work? Perhaps one way to refer to how the human–machine equation of value occurs is to look, for instance, into the machinehuman infrastructure sustaining Image Search Engines. The surrogate workers contracted at Amazon Mechanical Turk are, according to Atanasoski and Vora, “humans that perform the work of technologies that are claimed to replace the need for human workers” (2019, p 90). As the authors report, Amazon Mechanical Turk is a large-scale crowdsourcing software platform where human labor is contracted to become a service job to the machine. The racialized and gendered pool of global temporary workers for high data tasks therefore operates in lieu of an appropriate algorithm that can carry out the function. These become surrogate workers contracted to perform “artificial artificial intelligence,” that is a series of tasks that provides a given compensation in a given amount of time. For instance, “transcribe up to 35 s of media to text” is a precarious job that will pay between two to ten cents for the task. In short, human service enables the blind vision of machines to be matched to pre-existing categories that are performed by the Turkers so that machines can then perform the task themselves. In particular, filtering out randomness in generative adversarial pattern recognition allows for the vision machine to re-enter the ocular system of 13 exchange for which in the general equation of value entails how humans serve machines to serve humans. This autopoietic intelligence allow the equation of value between human and machine to double the extraction of value: the extracted labor of humans as service in making objects codable for machines only serves capital to extract more value from the human–machine equation of value. As much as the human Turkers become an agent of verification, dis-ambiguation, and de-volution, so too machines become agents of oppression that perpetuate the ocular epistemology of representation through the automated matching of images, concepts and objects. Elisa Giardina’s video installation "The Cleaning of Emotional Data" (2019) shows us how this Global South sociotechnical infrastructure that feeds the “artificial artificial intelligence” of surrogate humans has become internal to the dis-organic reproduction of global capital. In this installation, the blind image of computational media coincides with a recombinatory repository of data that the machine itself cannot see, and for which it needs humans to conform to a taxonomy of categories matching quantities of data. Giardina focuses on the Global South infrastructure of Turkers who get the service job of “cleaning” data by training machines to match images with emotions. These workers label, categorize, annotate and validate large amounts of data, enabling AI to recognize or order emotional patterns. Giardina herself worked remotely for several North American “human-in-the-loop” companies who provide “clean” datasets to train AI algorithms to detect emotions. Her own performance involves a taxonomization of emotions, the annotation of facial expressions and the recording of her own image to animate three-dimensional figures. While performing this work, some of the videos in which she recorded her emotional expressions were rejected by the companies she was servicing, because her facial expressions did not fully match the standardized list of affective categories that was given to her. However, as Giardina points out, it was not impossible to know whether this rejection originated from algorithmic protocols or, for example, from the consensus of fellow workers who supervize the service as they might have interpreted her facial expressions differently due to cultural contexts. "The Cleaning of Emotional Data", however, documents how the carrying out of facial expression does not easily fit a universal schema and instead the history of emotions questions the universal epistemology, where philosophical and psychological theories determine the meaning of facial expression and its mapping. Giardina’s installation makes the point that AI systems, which supposedly recognize and simulate human affects, base their algorithms on understandings of emotions that are universal, authentic and transparent. The cleaning of data from facial expressions that do not fit the universal schema of emotions is a human service job carried out by underpaid
P. 7
AI & SOCIETY workers that sustain what Atanasoski and Vora call the “surrogate effect” of the coming phase of artificial intelligence. In particular, these “technologies that erase human workers are designed to perform the surrogate effect for consumers, who consume the reassurance of their own humanity along with the service offered” (2019, p 91). Tech companies and governmental agencies use these human-verified data to develop software that identifies consumers’ moods or that recognizes facial expressions of potentially threatening people. However, the matching between facial expressions and the categorization of emotions entails the filtering out of noise or noisy emotions that persists as an internal tension within the universal model of visual representation, the ocular matching of concepts, objects and image. The correlation between knowledge and vision is reintroduced in the negative optics of machines that fail to recognize emotions—that is in the negative algebra that assigns a zero to a mismatched pattern—and becomes for Giardina a critical space to expose the internal paradox of capital extraction and of the critique of visuality today. In particular, Giardina’s work offers a response to the contemporary critique of the invisible image, according to which machine vision excludes the human from its systems of operations. Instead, her research points out that it is not that humans are excluded from the loop of the machine to machine communication, but instead that this new phase of planetary artificial intelligence and/or full automation fundamentally relies on a global re-organisation of racial and gender capital in terms of surrogate humanity. In addition, one can argue that within this re-organization there is also an intensification of the equation of value that includes the circuit of extraction where humans have to teach machines how to learn the ocularcentric matching of concepts, objects and images. Exploring the mono-logical economies of extraction across the Global South populations, Giardina’s work shows how the racialized and gendered precarious labor operates as the infrastructural components of artificial intelligent systems. As much as the infrastructural web of data cleaners, algorithms trainers, proof verifiers have become enfolded in the blind operations of databases through which machine learning algorithms, and image search engines, are taught to recognize patterns, the colonial epistemology of knowledge and vision continues to impart the equation of value between surrogate humans and machines. What is here excluded instead, as Giardina’s investigations show, is not the human, but what has always been less than human, non-human, and in-human coinciding with the precarious labor and the machine-like service of outsourced subsumed subjectivities absorbed in the operational image of machines. As much as the colonial capitalization of the human–machine alliance accelerates the global (and outsourced) modes of enslavement under the universal equation of value reified in the automated system of the decision today, so too humans are used to correct machine vision and to re-impart the visual categories of knowledge back into the system of acceleration of value. A spiraling extraction of the human–machine alliance under the universal model of technology is at play here. Giardina’s perspective about the intrinsic presence of human labor in the blind machine of techno-capital, however, is not simply a proposition aiming to correct the critique of the operational image by claiming that instead humans are included in the feedback loop of machine-tomachine communication. Importantly, one can suggest that her intervention brings to attention the limits of visual critique in the context of intelligent computation that relies on the ontological ground of knowledge and vision. In other words, by insisting upon the underlying universal equation of value extraction that connects (or, places in a dialectical mirroring relationship) blind images with outsourced, surrogate human labor, Giardina’s work pushes the critique of automated vision towards a radical engagement with the material conditions of exploitation or intensified extraction of value in human–machine labor today. Can these combinatoric modes of abstraction exceed the equation of value in the universal extension of capital? It is interesting how Giardina joins together the abstract lines of facial micro-expressions detected by the algorithms with untranslatable emotional vernacular from both Sicilian dialect and American English. In her collaboration with Michael Graham of Savant Studios, she weaves together computational and human language in a large-scale textile pieces, called Amiss Motifs, mapping unrecognizable emotional patterns with distorted, uneven, broken patterns to expose the overlapping of racialized surrogacy with the blindness of machines. In contrast to the critique of the invisible image that reifies a full automation that excludes the human, Giardina’s reflection on the surrogacy of the human–machine labor as the infrastructure of intelligent techno-capitalism instead points to how the negative optics—namely the cultural, affective and the aesthetic labor of the human–machine—breaks from the universality of technology and the dialectic between the visible and the invisible. In the next section, negative optics will be further discussed in the context of the epistemological capacities of automation of refusing the transcendental authority of representation. This will be explored by engaging with Laruelle’s argument for dark optics and immanent vision. In particular, the next section focuses on machine vision and discusses current attempts to reduce negative randomness in Generative Adversarial Networks. The attempt at correlating negative optics with negative randomness is to suggest that machine vision can be theorized away from the ocularcentric dyad of visible and invisible image. The residual negativity in computational randomness opens learning algorithms to 13
P. 8
AI & SOCIETY the indeterminacy of knowledge and allows a non-optical theorization of vision in computational systems. The last session or coda will discuss negative optics in relation to the negative dimension of blackness in the equation of value under racial techno-capital as a starting point for proposing an immanent epistemology. 2 Auto‑impressions In current research developments in machine vision, it is possible to contextualize negative optics in the field of dynamic geometry, such as mereo-topology (or the study of parts and the relation between parts and wholes), and artificial neural networks that are programmed to learn from patterns. Cognitive psychologist and computer scientist working at Google Brain, Geoff Hinton, claimed that the logic of neural networks on which machine vision is based is limited to conform to pre-established parameters (2017). In particular, Hinton addresses the need to re-design the procedural process by which algorithms can learn from each other in the neural network through what he calls “capsule network”—a form of AI that enables machines to understand the world with images without relying on existing parameters of vision. In his 2017 research papers, Hinton argued that capsule networks have not only led software learning to recognize handwritten digits, but they have also halved the error rate in pattern recognition of toys and cars. Since image recognition software is used generally, and thus not contextually, to recognize objects, the predictive patterning cannot learn to recognize the same object in different scenarios. This is why the surrogate human needs to verify and thus teach the machine that what it is seeing is actually the same object. Instead of relying on this external verification, according to Hilton, capsule neurons—small groups of crude virtual neurons—will track only parts of an object, for instance the cat’s ear and nose, as these are differently positioned in space. This smaller scale of algorithmic receptivity, according to Hinton, will enable a neural network to figure out the difference between scenarios by extracting more information from the mereo-topological relations between smaller parts of data. As the capsule network is made of smaller patterns set to recognize parts and break the continuity of an image into smaller units, Hinton designed a dynamic routing between capsules that trains these kinds of network. Capsule algorithms convert pixels fragments into vectors of recognized patterns and then apply a transformation matrix to these fragments to predict the parameters of larger fragments. In particular, the transformation matrix learns to encode the intrinsic spatial relation between a part and a whole, which results in the formation of an invariant viewpoint, a perspective or direction in recognition that aims to generate a novel 13 vision. According to Hinton, “capsule use neural activities that vary as a view point varies rather than eliminating variations” (2017, p 9). Instead of normalizing viewpoints according to methods such as the spatial transformer networks, or the automated filtering out of randomness, capsule networks simultaneously engage multiple transformations of different objects or object parts. This series of correlated activities are described in terms of “routing-by-agreement” (Hinton 2017, 2–3, p 6). As opposed to the optical filtering out of noise, this technique aims not to eliminate redundant neurons, but to use all non-averaged information to obtain a non-totalizing knowledge about the position of an entity in a region. The inclusion of micro-variations in machine vision run on top of CNNs, which instead are aimed at solving the disambiguation (or mismatch between concept, object and image) by playing out on the automated (or self-correcting patterns) of both reducing and increasing negative randomness in the neural network. It seems important to repeat here that CNNs are an instance of deep learning networks for machine vision. Convolution already applies a kernel to overlapping regions that shift around the image by eventually establishing a fully connected network through a one to one relationship of neurons across distinct layers. However, this multiple level of connection across layers also seems to incapacitate the system from learning new data. In other words, the optical matching on negative randomness seems to block the possibility of machine vision. When convolution leads to the overfitting of information as a one to one connection, the kernel risks to re-learn redundant data (that is the same data will be held in two places in the database). Convolution, therefore, delimits machine vision by saturating data memory and increasing computational costs for instance, but more importantly because its overfitting capacities lockdown algorithms into a pattern of connection between the same parts, without being able to learn from what cannot be known in advance. From this standpoint, since at each location in the image there is one instance of the type of entity that the capsule represents, the capsule model—as opposed to the convolutional matching of concepts and images—affords a form of distributed representation. This is inspired to the perceptual phenomena of crowding, where neighbor parts shed the direct perception of an object in movement. CapsNet architecture, therefore, grants not an algorithmic matching with existing data, but shows that algorithmic patterns can become predictive vectors of futurity (that is they look for what cannot be already seen) and not simply an automated vision of patterns recognition. Instead of eliminating variations to reach an average capacity for general recognition, predictive vectors follow the negative algebraic relation between layers, involving randomness and micro-temporal variations in the algorithmic process of compression. These
P. 9
AI & SOCIETY vectors of variation have become central to how machine learn beyond set parameters, and how negative randomness has become a source for machine vision entailing no transcendental recognition between concepts, objects and images. In other words, computational vision establishes a potential or hypothetical relation between non-correlated patterns that may correspond to images that did not yet exist. From this standpoint, predictive vectors are not simply the technical explanation for apophenia—namely how the invisible image makes decisional patterns—but are more than a set of probabilities based on what is already been recognized in the system: the negative algebra of what cannot be recognized pushes the system to construct counter-factual dimensions of images that do not correspond to the ocular equation of value between objects and concepts. It is as if the discretization of the network in increasingly smaller vectors of variations flips the architecture of matching inside out by exposing indeterminate dimensions to its organizational infrastructure. Instead of a self-reproduction of the master pattern across the layers of the network, the intensified discretization of networked parts also increases the volume of randomness within and amongst them. Computational vision can thus be defined in terms of the indeterminate series of variations that each time become cloned in each and any algorithmic patterns. Here algorithms do not simply register raw data to execute instructions but become sheer receptors or cloning of an indeterminate series of the real. One could argue that this negative optics of algorithms—that is unable to be a mirror of the world because it does not see the world but only a series of variable parts that entail a process of auto-impression of matter, whereby machine vision—becomes a medium for heretic inferences, for the elaboration of a heretic logic without ocular-logos. In other words, a dynamic bootstrapping between algorithmic patterns and vectorial randomness suggests that machine vision is embedded in a series of temporal variations that become finite inferential hypothesis through the auto-generation of material images. As machines become auto-impressions of an infinite real they also generate and envision new patterns: the dynamic movement between increasingly large and increasingly small scales of predictive vectors turns machine vision into a heretic auto-impression of computational matter. This process of auto-impression of matter without ocularlogos is opposite to what Laruelle calls “algorithmic transparency” which rather takes the technical image (especially understood in photography or photographic philosophy) as that which measures the correlation between premises and results in terms of effects (or the efficient causality of enumeration of given ends), and thus presupposes homogeneous matching between the transcendental order of concepts and objects (Laruelle 2013). Borrowing from Laruelle, one can argue that the ocularcentricism of philosophy coincides with the decisional power imparted on the real by the transcendental mirroring of knowledge and vision (Laruelle 2013 By challenging the optics of philosophical decision, Laruelle’s non-standard philosophy or non-philosophy makes the argument for an abstract theory of photography, “absolutely nonworldly and non-perceptual” (date: 8). In particular, Laurelle asks: “to what extent is photography not an activity, for example, of a kind with Artificial Intelligence (AI)—an attempt at the technological simulation not of the World, in its objective reality, in its-philosophico-cultural reality, but of science and of the reality that science can describe, naively in the last instance?” (2013, p 10). What Laruelle is insisting upon is to refuse the universal paradigm of perception grounding from the standpoint of being-in-the-world. Instead, his non-philosophy takes photography as a technique as a form of knowledge that introduces science in the condition of existence of perception and of the world (2013, p 11). Laruelle’s non-philosophy takes the critique of the ocurlacentric vision to another point, further delving into how negative optics refuses transcendental decisionism and at once lays out the possibility of elaborating a non-optical dimension of knowing, a non-ocularlogos of knowledge. For Laruelle, the condition of knowing depends not on the point of view of philosophy, but it is rather to be found in what he calls “the stance” of, borrowing from Deleuze and Guattari, “a body without organs” (1989). Instead of a self-determining reflection of the world, vision becomes un-objectivating, implying not a position, a decision, but an auto-impression that is first of all a real “undivided experience, lived as non-positional self-vision force” (2013, p 13). If, according to Laruelle, photographical decision corresponds to the law of sufficient reason that reflects the real according to a circular self-expression of truth, non-photography (as a non-philosophy) therefore coincides with what he calls “fractal algorithms,” because they have a degree zero of self-reflection. Following this argument, one can suggest that the algorithm is not a medium programmed to reveal the world or even less to self-regulate the human perception of the world. In other words, it is not a prosthetic tool that ensures constant adaptive feedback. Instead, the fractal algorithm is increasingly partial and in this fashion clones its own real image, namely of cloned image without original or copy, in terms of a spatial surface that extends (or infinitely fractalizes) forever without uncovering any pristine form behind it. Laruelle explains that what appears in the photo as an object drawn from the transcendence of the world must be distinguished from the “photographic apparition” (2013, p 18). By radicalizing the Husserlian distinction between the photographed phenomenon, corresponding to what photography can manifest and thus to the manner in which it manifests the world, and the photographed object, namely the representation of the world, Laruelle argues that 13
P. 10
AI & SOCIETY the non-photographic vision (or negative-optics) is a parallel process to the world, and as such it is not in the world (2013, p 25). This is not the field of transcendental knowledge but defends real immanence in vision, or what Laruelle calls “vision-force” that is immediately given or the “in-itself” of the image. Thinking with Laruelle’s non-photographic argument for vision, it is possible to push further the heretic account of machine vision and thus refusing turning machines into the onto-epistemological mirror of transcendental vision. One could argue that as much as negative optics neither represents the world nor remains an invisible unreflexive automatism of the world, it crowds the parallel space of photographic apparition. This is not a representation but an auto-impression of an image’s own aesthetics that starts from the negative fractality of algorithmic micro vectors that give the effect of what Laruelle calls “generalized fractality” (2013, p 78). Instead of serving as an instrument for shedding light onto the world, the camera-machine clones (equate without equivalence) the underworld of dark optics as fractal patterns of an indelible generic intelligence that turns on its own head the ontic limits imparted on science by philosophical decision. As Alexander Galloway puts it: “[i]nstead of mere ontic darkness, generic being achieves an ontological darkness, and hence beckons toward the kind of crypto-ontology of pure blackness evident in Laruelle” (2014, p 77). If one were to continue along this line of argument, it is possible to suggest that a crypto-ontology is what can define the negative optics that is foreclosed to a transcendental being. On the other hand, however, negative optics is precisely the algorithmic randomness that demarcates ontological darkness in the machine in terms of an auto-impression of machine visions that signal the apparition of the image itself. These are infinite reflections without mirror, “unique each time but capable of an infinite power ceaselessly to secrete multiple identities” (Laruelle, 2010, p 82). This multiplicity of darkness, one needs to emphasise, coincides not with substantial forms, but with non-consistent phenomena that entail a certain, non-optical automatism in exposing, in Laruelle terms, the “hyperphenomenology of the real” (2010 , p 95). In particular, as Laruelle specifies: “There is a ‘phenomenological’ automatism or blinding that culminates in the photographic eviction of the logos – of philosophy itself – in favour of a pure irreflexive manifestation of the phenomenon-withoutlogos” (2010, p 95). The scope here is not simply to unmask the supremacy of self-determining philosophy in the name of absolute irreflexivity but to rather start from this un-mirroring image as the stance of negative or dark optics. Here the automated image become the medium of auto-impressions as the multiplicity of darkness pulls through the phenomenon-without logos in machine vision. The automated image exposes the fractal 13 consciousness of a machine knowledge that turns the ontic limit of epistemology into singularities or non-axiomatic automatisms. This is the artificial image that stays with and pushes further the negativity of the human, the non-human, the less-than-human, occupying the stance of a “[s]tranger in flesh and blood” (2010, p 103). Laruelle’s arguments for non-philosophy radically defies the façade of pretentiousness of Western metaphysics for which the real can be surgically cleaned from blackness, which remains locked in the form of the Other that must follow the image of Man, once it has been emptied out and turned into a soulless machine. For non-photography instead only the other has to be replaced with a “logic of auto-impression” or phenomenological automatism where the medium becomes a negative machine of a black universe (Galloway 2014, p 191). From this standpoint, if we are to follow Laruelle’s proposition for non-axiomatic photography, the theorization of computational vision may have to start with algorithmic fractality, namely the possibility of a computational autoimpression of the real, exposing the alliance amongst lessthan-humans in a field of immanent vision. The Laruellian “Vision-in-One” is a stance that does not predetermine but rather becomes determined by the real in the real’s “last instance.” As much as this determination is a clone of the real, but not the real itself, one can argue that machine vision entails a non-relationality with the world, whose negative auto-impressions are infinite singularizations of blackness beyond the optical value of representation. That non-photography insists on the cloning of the black universe however is not another way to propose a critique of automation. Instead, this is above all a practice of refusing, hacking, alienating the transcendental decision for which machines as blind can only learn to represent what is already given to them by the ontic limit of knowledge. Similarly, non-philosophy is not simply an invitation for imagining alternative ways to reinstate philosophy and expanding its ontological ground of transcendental decision. Instead this is a heretic project that necessitates trans-collective elaboration of fictional automations that start with the auto-impression of blackness as proliferating each and anytime outside the onto-epistemological cosmogonies of the self and the other. For instance, one can start asking why techno-capital decisionism is still taken as the onto-epistemological law that locks the critique of computation into the perpetual mirroring of vision and knowledge, humans and machines. The fractality of the algorithms can rather contribute to engage randomness in process imaging as machine singularizations of the real. In other words, machine vision can become a medium for the auto-impression of darkness in the last instance: namely as machines learn to learn from negative optics. The negative materiality of the computational image does not only show that knowledge can be divorced from
P. 11
AI & SOCIETY transcendental ocularcentrism, but also that the invisible image of machines is part of alien epistemologies that defy the equation of value with the infinities of 0 s. The fractality of algorithms is part of the expansion of heretic epistemologies where alliances between the configurations of the in-human demonstrate how the equation of value can be not only challenged, but turned inside out from the stance of infinities—the auto-impression of blackness. In the following coda, this argument for the algorithmic fractality of infinities that accounts for the invisible images of machines outside ocularcentric critique requires further collective elaborations of how negative optics can become part of a revolutionary practice of machine thinking (that indeed overturns the critique of ocularcentrism within technocapitalism). This coda, in particular, turns to the elaboration of negative infinities in Ferreira da Silva’s articulation of blackness as indeterminate matter at the limit of modern thought, whereby matter without form is above all “matter beyond the equation of value” (2017). 3 Coda on images without value Drawing on quantum mechanics, for which indeterminate results in determining the perception of reality point to the necessity of moving beyond the ontic limits of science, Ferreira da Silva discusses this indeterminacy or the Thing as the referent of blackness, another mode of existing at the limit of modern thought. As she puts it: “deployed as method, blackness fractures the glassy walls of universality understood as formal determination (2017, p 11). She proposes an experiment in articulating the “equation of value” as a self-determining formalism based upon and through the violence against the thing/blackness. Ferreira da Silva proposes to carry out this algebraic experiment not simply to unveil the formal determination of the matter at the core of the equation of value, but one could argue, to rather address the “auto-impression” of blackness minus the form, or matter without the universality of value. To do so, Ferreira da Silva proposes an ethical re-articulation of blackness that does not follow the over-determinant image of Man whose program overfits – that is overrepresents—all modes of being human nor, on the other hand, aims to make a claim for an absolute outside that separates the Modern/European/ white human from the non-modern/non-European/black non-human (2017, 04/11). Ferreira da Silva argues that this ethical experimentation starting from matter beyond the equation of value deserves further investigation about how determinacy, together with separability and sequentiality have sustained modern thought and the construction of an ethical matrix in which the indifference with which racist violence is met, is itself rather become constitutive of a (common and public) moral stance (2017, p. 04/11). She traces how this matrix is entailed in the modern elaboration of causal efficiency and its operative grounding of equivalence mathematically bounding together ethical, economic and juridical formations (2017, p. 04/11). The modern Kantian world becomes a way to bring together formal and efficient causality in the self-determination of the limits of scientific knowledge—precisely concerning what can already be accessed by the senses (the empirical experience of which science provides the tools for extending universal measure). As value becomes universal and moves across scales, the object (thing/matter) is unified by its formal qualities which in turn are the effects of judgements (and thus transcendental concepts) derived from the measurement and classification of objects (that is by the ontic limits of science). Here the difference is granted by the decisional position or transcendental operator that already knows the object. Ferreira da Silva explains that within this transcendental field of value, blackness as a category of racial difference “occludes the total violence necessary for this expropriation [namely the colonial expropriation], a violence that was authorized by modern juridical forms – namely, colonial domination (conquest, displacement, and settlement) and property (enslavement) (2017, p 08/11). However, Ferreira da Silva’s invocation of the Thing in quantum mechanics, as much as it claims for an autonomy of matter beyond form or universal equivalence of value, also offers a way to unsettle the universal matrix of ethics. The Thing challenges the ontic limit of knowledge imposed on and through science by transcendental philosophy, which saw the extension of the formal into efficient cause driving modern colonial epistemology. In particular, Ferreira da Silva’s ethical experiment offers us a proof of the “equation of value” for which blackness as nothing—that is zero value or infinity—has the creative capacity to unsettle and hack Modern Western onto-epistemology of vision and knowledge. According to Ferreira da Silva, zero is to be taken not as a contradiction, as established by the dialectic of presence and absence, but itself as a signal of the autonomy (or in Laruelle’s term unilaterality of) negation and of the negative. Instead of being invisible, blackness, as matter without form, brings forward the nullification of the ocularcentric field of vision. But how to demonstrate that such a nullification is not simply another manifestation of the contradiction that rather confirms the norms of the transcendental? This is a question that must accompany the theorization of the negative optics of computational vision because it must bear the challenge against the ontic limits of knowledge that continue to be inscribed within the performance of science of information, whereby the potentiality of machine epistemology is constantly resubsumed by transcendental decisionism and representational metaphysics. For Ferreira da Silva, it is a question of re-articulating the mathematical mapping of the matrix 13
P. 12
AI & SOCIETY of ethics in terms of a radical engagement with zero as a value in itself, rather than self-positing a critique that reveals contradictions in the transcendental equation of value. In as much as the result of Ferreira da Silva’s experiment with the equation of value points to the real dimension of the undeterminable (namely value without form, namely neither life nor not-life), it also has zero value because it exists itself without form. In equating blackness with zero or infinity, Ferreira da Silva proposes a radical praxis as a refusal of dialectics founded on the decisional principle of philosophy, whose complains against (and rectification of) the invisibility of the automated image, are there only to re-produce its ocular-logos of recognition (2017). From this standpoint, it is possible to conclude that the dissolution of determinacy and its transcendental decisionism requires not a skeptic critique against computational vision but above all a creative elaboration of the non-ontic science of vision machine that rather enables the possibility of overturning dialectics and its critique. The argument for negative optics, therefore, does not only concern the human–machine relationship but more importantly how this latter, under technocapitalism, can rather be turned inside out to defy the onto-epistemological violence of ocular vision through and with the negation/negativity of infinities, with blackness as matter without form, auto-impressions of the real without the given, auto-imaging minus the self-positing subject. From this standpoint, a praxis of refusal stays with the trouble of dark optics, plunging within multi-layering opacities, fractalizing blackness in artificial visions. From this standpoint and against the universal model of technology (the techno-logical universalism of ontic knowledge), it seems crucial to continue to elaborate how machine vision is equivalent not only to the perceptual gaze but also to the operational image that has sustained, since modernity, the colonial operative epistemology and its matrix of ethics. Ferreira da Silva’s reversed equation of value is a praxis of refusal that does not only reveal the violence of self-determination but also invites us to stay with the negativity of algebraic relations for which nothing is a signal of infinity running against the metaphysical equivalence of knowledge and vision and the capital equation of humans and machines. If we were to invert the equation of value in computational processing one may need to start from the indeterminacy of compression, the negative algorithm of what cannot be known in advance, the negative randomness that condemns the computation to remain incomplete. Since negative optics implies no image-form, so too no universal formalism can grant the outcome of the algorithmic compression of randomness. It is, therefore, possible to argue that negative optics is an index of immanence—namely no transcendental value can explain the auto-impression of images in the algorithmic discretization of increasingly smaller patterns of recognition. Here the networked relation between image, 13 concept and object is turned upside down as much as the auto-impression of images exceeds the operative correction of images that capital imparts on the human–machine alliance. Instead of the ocular representation of the world, the increasing discretization of the network increases the volume of randomness as much as the ANN affords the ingression of indeterminate variations within algorithmic compression. In other words, the machine vision also coincides with what cannot be explained, programmed, represented beforehand by concepts programmed in the machine. This is why algorithmic operations are not simply prescribed perceptions but are auto-impressions of the human–machine condition each and any time multiplying the field of auto-apparition of blackness without form in the last instance. To conclude, it is central to this argument that as much as the negative optics of computational processing speaks of the challenge of reinventing epistemology outside the knowledge-vision correlation, it also requires us to unsettle the colonial roots of modern epistemology for a space of thought for asymmetric auto-impressions of the human–machine alliance. This is a collective and transversal effort in thinking with machines that starts from acknowledging the inhuman condition, the negativity of value, zero value in the human–machine alliance. This is also to say that if the modern universality of technology continues to become reified in machine visions, so too the argument for practices (all forms of practices) starting from the zero value of auto-impression must continue to unpack the heretic versioning of negativity in the technopolitical reconstruction of epistemologies. Compliance with ethical standards Conflict of interest Nothing to declare. References Atanasoski N, Vora K (2019) Surrogate humanity. Race, Robots and technological futures (perverse modernities: a series edited by Jack Halberstam and Lisa Lowe). Duke University Press, Durham and London Ferreira da Silva D (2017) “1 (life) ÷ 0 (blackness) = ∞ − ∞ or ∞ / ∞: On Matter Beyond the Equation of Value.” e-flux 79, February. https​://www.e-flux.com/journ​al/79/94686​/1-life-0-black​ness-oron-matte​r-beyon​d-the-equat​ion-of-value​/ Galloway AR (2014) Laruelle: against the digital. University of Minnesota Press, Minneapolis Giardina E (2019) The cleaning of emotional data. Aksioma—Institute for Contemporary Art, Ljubljana. https​://aksio​ma.org/clean​ing. emoti​onal.data/ Hinton G et al (2017) Dynamic routing between capsule. In: 31st conference on neural information processing systems, NIPS 2017, Long Beach, CA. https​://arxiv​.org/abs/1710.09829​. Accessed 22 Apr 2020 Laruelle F (2010) The concept of non-photography. Urbanomic, Falmouth
P. 13
AI & SOCIETY Laruelle F (2013) The transcendental computer: a non-philosophical utopia. Trans. Taylor Adkins and Chris Eby, Speculative Heresy, August 26. https​://specu​lativ​ehere​sy.wordp​ress.com/2013/08/26/ trans​latio​n-of-larue​lles-the-trans​cende​ntalc​omput​er-a-non-philo​ sophi​cal-utopi​a/. Lowe L (2015) The intimacies of four continents. Duke University Press, Durham, NC Paglen T (2014) The operational image, e-flux. November. https​:// www.e-flux.com/journ​al/59/61130​/opera​tiona​l-image​s/ Paglen T (2016) Invisible images (your pictures are looking at you). The new enquiry. December. https:​ //thenew ​ inqui​ ry.com/invisi​ bleimage​s-your-pictu​res-are-looki​ng-at-you/ Steyerl H (2015) A sea of data: apophenia and pattern (mis-)recognition. e-flux. April. https​://www.e-flux.com/journ​al/72/60480​ /a-sea-of-data-apoph​enia-and-patte​rn-mis-recog​nitio​n/ Virilio P (1994) The vision machine. Indiana University Press, Bloomington, IN Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 13