All Tomorrow’s Pictures
January 27, 2015
Art celebrates the unique and immediate experience of the “what it is like” to perceive the art object. It has become accepted, at least since Greenberg, that the refinement of Art to a subjective encounter between the viewer and the art object is, in-itself, the defining characteristic of artistic experience. With the modifications of Duchamp’s gesture, which sloughed off the categorical limits Greenberg had once restricted Art, it is now solely this irreducible immediacy of perceptual experience (so long as it is properly framed in an artistic context) which is signals the reception of an experience as Art. For Art, this experience is tied to a fundamental material and biological limits of the human body. What the human is remains essentialized in its discourse, so long as we have no means of conceptually or materially superseding this horizon. However, while our conception of the aesthetic remains tied to the givens of human experience, we are now beginning to understand how this given experience is constructed, and perhaps even develop the capacity to reconstruct it. Machine vision, and its nascent algorithmic construction lays bare the functional and mechanistic principles which underline the transcendental horizon, perhaps heralding a new aesthetic.
Under the skin
In Beyond Good and Evil, Nietzsche (as was characteristic of his rhetoric) made a rather caustic remark on Kant’s synthetic judgements a priori:
“There came a time when people rubbed their foreheads. People are still rubbing them today. They had dreamed: first and foremost – the old Kant. ‘By means of a faculty,’ he had said, or at least meant. But is that an answer? An explanation? Or is it not rather a repetition of the question?” (Nietzsche, Beyond Good and Evil)
Questions of how it was possible that we could form synthetic judgements became especially pertinent for the empiricists, who attempted to show how beliefs about the world could be known and made tractable upon our conception of it. For, if we accept an internalist position that we have access to only our own beliefs about the world, and not the world itself, how might what we seem to know about the world accord with it? Following in the footsteps of the Cartesian tradition, our surest beliefs regarding the world were thought to be developed through some directly known experience, which in its immediacy and observability provided the most basic foundation for our empirical judgements. Our faculties were thus presumed to be in harmony with nature, causing us to hold the most basic facts regarding existence, such as that of our own existence.
However, this picture elides the distinction between the metaphysical intuition of the primacy of causation in the material realm, and the epistemological role of justification in the mental realm. For, if our sensual experience causes us to have some elementary beliefs, this only pushes back the problem of how it is that something material can come to have the form of something that is mental. Our beliefs take primarily propositional form, and thus can be analytically dissembled, while material things are not propositional in form. A cause cannot be a belief, since it lacks the proper epistemological status.
The philosopher Wilfrid Sellars proposed the distinction between the manifest and scientific images to help frame and overcome certain perennial problems in epistemology and ontology. While the manifest image indexed the developed image of humanity in the world as it sees itself, the scientific image presents a picture of humanity as it is in the world. While from a manifest perspective, certain epistemological and ontological problems are undecidable, because their solutions rest beyond what is fundamentally empirically observable within the manifest image, the theory laden conception of the world that is regulated by science should be capable of extending the capacities of human thought and action beyond our parochial understanding. It is clear that science outlines particular properties of the world, such as Einstein’s Special Relativity, that are not party to our manifest observables (and even seem counter-intuitive to them), yet nonetheless allow us to predict and explain aspects of the world. However, Sellars did not seek to totally eliminate the manifest image, but rather to bring it into accord with the scientific image, such that we would be compelled to update our normative conceptions of what the human is and could be along the lines which science commits us to.
As Nietzsche pointed out, the Kantian manifold is underlined by processes and capacities for which it itself has no explanation. For Sellars the ultimate explanation for this trick of psychology — this capacity of the mind — need not be explained in terms of a philosophy, but is the remit of a scientific theory which will outline how it is that minds have the functions that they do. Sellars intended to thoroughly naturalize the capacities of the mind, in a causally reducible manner. This is not to say, however that persons are themselves logically reducible. His position is, in a Kantian vein, that the normative framework of our concepts is compatible with a natural causal framework, and that one cannot be reduced to the other for the simple fact that concept of persons is not the same as the concept of the bundles of capacities which make up a person.
Do Androids dream of electric sheep?
Sellars recognized the distinction between information from semantic content and information relayed from physical systems. As I have pointed out above: while language operates according to a normative logic, empirical data derives from the causal structure of reality – one is a conceptual understanding of the world, while the other is non-conceptual. There is, therefore, a problem as to how these two types of information connect with one another to form a picture of the world that bears some resemblance to it as it is. Sellars identifies the ability to recognize isomorphisms in linguistic systems with uniformities in the empirical world as the hook upon which language is capable of latching into the world and getting a grip on the real. Sellars uses the relations of picturing and signifying to describe these two systems of knowing, telling the story of a robot which uses sensors to observe its environment and recording descriptive sentences on a tape for every new feature of the environment it comes across. Over time, the sentences of the tape will produce a more and more complete description of the robot’s environment – producing an account which signifies and isomorphically corresponds with the space in which the robot is situated.
This analogy helps to characterize human thought in terms of both a functionally normative realm of language and a functionally causal world of sensual perception. While the exact mechanisms are never described, it is useful to consider the difficulties involved through the lens of AI development. Deciphering image data in natural language terms has been a particularly hard problem for AI researchers since it was first proposed. In 1966, the MIT computer professor, Seymour Papert, proposed as a summer research project for his students machine vision, including the problems of pattern recognition, field ground relations, and object and region description. Forty years on, we are just beginning to make significant progress with these problems. For example, both Stanford and Google researchers have recently advanced the natural language descriptive abilities of machines. These “deep learning” machines apply complex, hierarchically situated models to image recognition to weight and detect patterns and figure/ground relationships in images. These generalizing capacities have been cross-referenced and vectored with other algorithms whose function is to detect patterns in languages and translate from one language to another. Rather than substitute another language for the pattern detection system, however, researchers are now supplementing the pattern detection systems for images in place of the dataset for another language. Models then, which analyze the functional role of a term in a coherent language system then, are correlated with models which generalize patterns in images. This technique is remarkably efficient in automatically generating natural language descriptions of images that are readable to a human being. Most interestingly, these developments seem to be anticipated by Sellars’s division of normative and empirical pattern detecting systems, and their ability to map onto one another to produce semantic statements. Both normative and empirical faculties are slowly being approximated through an understanding of sub-processes which simulate their functional roles. That higher level processes, such as language and sensual experience might be decomposed into lower level processes which are scientifically descriptable seems more immanently achievable than it was in Sellars’s time.
This research does not yet approach the level of human intelligence, especially since these machines lack the general conditional concepts of coherence which would be necessary for them to truly be ‘knowers’ in Sellars’s sense. The machine only maps the conditions which give language coherence, but does not seem capable of reflexively explicating and applying said concepts. To be a ‘knower’ means having a grasp of the general meta-linguistic rules which govern a language. For instance, to even have the capacity to reliably state that I am having an experience of “a red triangular thing” over there, requires that I have what Sellars calls a “battery” of concepts, including “red” and “triangular”, which only become distinguishable on the basis of coherent network of concepts. That is, for me to say that something is “red”, I must be able to distinguish it from other color concepts which are not “red”, such as “orange” and “pink”, and it is only in relation to these other concepts, and my ability to reliably indicate the distinction that I can be said to have a command of the concept “red”.
“Meaning” for Sellars, then, is only meaningful as it regards languages. This is to say that the term “red” acquires its contentfulness not simply because it refers to our non-linguistic sense of some stimuli that signals “red”, but because the term is situated in relation to other linguistic tokens in the system which, on a meta-linguistic level provide rules for its deployment in relation to the other terms of the system, and meaning is asserted by following permissible patterns within the rules of the language game. This is what is known as a “coherence” picture of languages, for to have a sense of one concept requires having a coherent set of concepts within which that concept is situated. Sellars adds to this the stipulation that these terms must also be deployed in conjunction with the uniformities of nature to be appropriately expressed.
A major objection to this picture of the reliability of a knower is the question of infralinguals — that is, how could it be that animals, babies, and now some machines seem to be able to reliably indicate things such as colors if they lack the linguistic capacities necessary for making these distinctions? One possibility is that symbolic systems necessary to be fully accredited as a knower, as is seen in the case of our image detecting machines, develop piecemeal, if at all, in some observers. That a child, or a machine, may know some of the rules of language, at least so as to be able to report the presence of a color or pattern, but not know and be able to manipulate the meta-linguistic rules which govern the relationship of the symbolic term seems plausible. In fact, most language speakers do not have to critically reflect upon these conditions unless pressed to make explicit their justifications for saying this is “red” rather than this is “orange”.
“[It] is in vain that we say what we see; what we see never resides in what we say. And it is in vain that we attempt to show, by the use of images, metaphors, or similes, what we are saying; the space where they achieve their splendour is not that deployed by our eyes but that defined by the sequential elements of syntax.” (Foucault, The Order of Things)
Many in the Artworld would take issue with this framing of our experience and its explicit reliance on linguistic concepts. They might say Art is not a propositional language — that is, something that we understand linguistically, and so, is not subject to this perspective. Yet, as we have pointed out, direct sensory experience of the kind that the empiricists wished to make foundational of our knowledge, cannot play the epistemic role necessary for knowledge. Furthermore, differentiating the qualities of our sensual experience from a pure field of sensual data requires some method of abstracting and symbolically locating that experience in the form of a pattern to make any sense of it. Nonetheless, there remains a qualitative difference between the experience of some sensual data and its explication in linguistic terms.
I have discussed how it is that we come to have concepts of our experiences, and noted that we cannot have a given experience without some conceptual content. For, to even be able to distinguish our experience, we must be able to explicate it in normative terms. The idea that Art, therefore, must excavate pure perception excepted from conceptual content cannot hold. There is no pure perception as Greenberg wished to nail it to the wall. However, that our experience is still limited materially and biologically by the particular processes and subprocesses of the manifold means that we cannot yet begin to step outside of horizon of given human aesthetic experience except in purely abstract conceptual terms. That is, so far as the theory laden discourse of science provides a normative framework for the world as it is, we have access to that picture, but not in a manifestly observable experience. Here Art may cling to its self-confined parochial nature by insisting upon the qualitative difference between the successive attempts to enumerate the observable variations of manifest experience and the character of that experience itself. It would therefore be Art’s task to horizontally expand our knowledge of the infinite variation available under those given constraints.
Beyond the garden wall
I believe the most immediately available exit from the horizons of a manifest Art is, following Sellars, to take a synoptic approach. That is, to make the critical interpretive task of Art explicit in normative terms, rather than relying on vague conceptions of given experience in artistic contexts. This would require a resuscitation of critical and theoretical institutions which Contemporary Art has sloughed off, but without the concerted effort to rebuild the discourse of Art’s positive conceptual dimensions, we wallow in market vulgarities.
The more speculative approach is to consider the unfettering of the material constraints upon thought. I have briefly discussed infralinguals as those beings which have sub-sapient intelligence, but let us consider what I will call supralinguals. The supralingual would have all of the capacities of a knower, with full command of the meta-linguistic rules of language, but also the capacity to explicate and cohere the very abstract models which undergird its own conditions. Consider the machine vision algorithms we discussed earlier. With the full capacity of general intelligence, speculative future machines may have immediate access to conceptual propositions whose level of abstraction is beyond that of standard human cognition. Furthermore, with the potential to reconfigure the functional capacities of their own conditions of knowledge, the types of input and efficiency of processing may be mutable in ways we could not imagine. So, in the case of visual processing, they may have the capacity to intuit an undifferentiated field of image information into a cohesive manifold which would be available for us with great difficulty and only under the mediation of linguistic abstraction. Of the “what it is like” for this peculiar inhuman aesthetics, we see through a scanner, darkly.