https://d2ng68cf36kxsk.cloudfront.net/store/public/post/1186/post_images/Slide357.JPG">

The failure of the Kodak company to recognize the significance of this invention became the failure of this image manufacturing giant, which after a long and painful decline fell into bankruptcy in 2012. The first commercially released digital camera was the Fuji DS-1P of 1988.⁴¹⁶

Digital cameras have developed over the decades, not because there has been any change in the principles of digital image capture since 1969, but as a consequence of increases in the sheer recording and calculating capacities of computer chips. According to Moore’s law, every two years, twice as many transistors can be “crammed into” an integrated circuit.⁴¹⁷ This prognostication has proved to be more or less correct in the half century after he wrote his 1965 article.

There are millions of pixels^§MS1.2.2e (megapixels) in each image taken on a phone, tens of millions in higher-end digital cameras. Considering that each pixel is represented by up to 24 bits of information about the light at a particular coordinate, a typical digital image may be decomposed into hundreds of millions of elemental zeros and ones before it is recomposed. Much of this is just alphanumeric naming of the quality of one pixel after another, column by column, row by row, in order to reproduce the same name as a color in another place. The algorithms in the first instance are simply to record and then reconstruct the array.

Then comes compression, where the algorithms that drive the near-universal file JPEG format remove up to 90 percent of the data because it is minimally useful to the reconstruction of the image. Several hundred pixels close together that are more or less the same color do not have to be named in their 24-bit specificity several hundred times. Instead, there can just be a command, “repeat however-many hundred times,” which will be mostly good enough even though in the capture they are likely slightly different.

Developed in 1986 and since then maintained by the Joint Photographic Experts Group,⁴¹⁸ we encounter here yet another place for the establishment of the universalizing frame of meaning that is digital modernity. The group’s members are nominated by government standards bodies, but they mostly come from businesses that have an interest in digital imaging. Attached to digital image files are Exif data – Exchangeable image file format, another international standard⁴¹⁹ – adding metadata about geolocation, date, device, and camera settings. Since 2015, JPEG has been complemented by the tighter compression of the High Efficiency Image File format (HEIF or HEIC) developed by the Moving Pictures Expert Group, interoperable with JPEG.

Cameras have always in some ways and to varying extents lied,^§MS2.2.2a but algorithmic readjustment of the light entering the lens in digital cameras helps them to lie more. High dynamic range (HDR) creates an evened-out luminosity across the image, reducing contrasts to bring some things out of the shadow and to make visible other things that would otherwise have been blinded in comparative brightness. Eyes do not work this way, taking time to adjust to significant variations in the intensity of light. In film photography these are areas of under- and over-exposure. Actually, a more accurate name for HDR would be unrealistically flattened dynamic range. As a consequence, the digital image becomes more rigorously space-bound in its planar simultaneity.

Near-infinite depth of field in digitally processed images also brings three-dimensional reality onto an absolutely two-dimensional plane, again in a way unlike the eyes or the blurring of unfocused backgrounds and foregrounds in film photography. The trompe l’oeuil of the digital photograph is in this respect more like painting than film photography. In this way too, digital photographs accentuate the planar nature of visualization in “unrealistic” ways. The more “realistic” they become, the greater the distance of image from the realities of seeing. However, in a sense, this is a late modernity where more is the same – in this case, a culture of visual distortion that creates illusions of realism is as old as linear perspective.^§MS2.2c

In digital photography, from the start we have mostly had a lot of numbers do a lot of naming for a lot of things, transpositions from image elements, to alphanumeric names for the qualities of light, to image element. The calculations are of a laboriously cumbersome kind that no human could practicably undertake – this is what Turing was referring to when he spoke of machine intelligence.^§MS1.3.2b We could if we wished do this kind of simple calculation manually with a small array of pixels, but without mechanical support, never many.

Then comes the much-hyped “artificial intelligence.”⁴²⁰ Digital imaging is a good place to explore the general nature of these particular processes of transposition via calculation.

Apple tells us that their phones use “machine learning” in a “Neural Engine … to analyze data from the camera sensor, quickly distinguishing faces in the frame.” This, in a revealing mixture of its metaphors, Apple calls “facial landmarking,” and “depth mapping.”⁴²¹

Before we reach the depths of “neural nets,” we must pass by machine learning. There are two kinds of machine learning.⁴²² The first, supervised machine learning, is where humans train the machine. Lots of human trainers tag this part of an image as a face, and another part as sky. When this combination of pixels comes up again, the computer can “say,” this is a “sky,” and this a “face.” Then, if I tag a particular face with a person’s name a few times on social media or in image storage software, after a while the computer will suggest their name if it “thinks” the same person might be in another photograph. This is a second kind of machine learning that is called unsupervised, where the computer notices repeated patterns and asks the user whether there is a label that might be applied.

So, the machine learns – but not because it ever really knows faces in general or this face in particular. All it knows is zeros and ones, for the purposes of naming and calculating, two vectors of transposition in computing that we discussed earlier.^§MS1.3.2 To this, we now add a third vector, rendering, or in the case of images putting the zeros and ones back together into a picture.

These are the text ⬄ image transpositions, mechanized from: view image > tag text > calculated pixel comparison > view image with suggested text tag. The stuff in the middle is intelligent for the number of calculations and their speed, but crushingly dull in the quality of each calculation. The limits of the meaning in this case are the limits of its transposability.

No two pictures are the same, so our theory of design explains.^§MS0.4.1 This means that the ostensibly intelligent calculations in the middle can never be “correct.” They can only be probabilistic – likely this is a person, then likely it is this person. The machine becomes smarter each time a user confirms or denies a suggestion. But the elemental calculations can never become smarter. The intelligence is in the “bigness” of data, and not (so much) the sophistication of the probabilistic statistics.

Neural nets involve multiple layers of pattern recognition, finding patterns in the numbers, then patterns within the patterns, again and again until they become fuzzier, then working their way back from highly abstracted fuzziness to other instances of the pattern which are probabilistically similar.⁴²³ This can’t work without a massive amount of data. And it is only neural by analogy, as if the brain can be reductively reduced to 0/1 calculations, which it can’t. Its roots are a statistical technique less than inspiringly called “back-propagating errors” in the 1986 paper by its creators.⁴²⁴

Soon, we are now told, will come quantum computing, though this may be just another kind of probabilistic statistics, still built on the mechanical capacity of electricity to be on and off.

The only artificiality in artificial intelligence is in the mindless granularity of the datapoints and the scale of the calculations. The rest is human intelligence, transposing the meanings of their images into text, and the computer repeating our own words back to us when a similar array of pixels turns up.

Or, to say this in the more general terms of our three vectors of transposition, in from medium to calculation and back again to medium. Naming: the computer decomposes and names mostly miniscule things in unspeakable ways. Calculation: the computer counts them in ways that could not sensibly or practicably done by hand. Rendering: the computer gives the results of the calculations back as image (and its variant, text) and sound (and its variant, speech). This is how AI works, no matter how fancy the statistics. It is a process of functional transposition into and out of quantities, and multimodal transposition, such as the text ⬄ image dependencies that underpin machine learning. AI’s power is in the grammar of its transpositions, not its mathematics.

---

⁴¹³ W.S. Boyle and G.E. Smith, 1970, “Charge Coupled Semiconductor Devices,” The Bell System Technical Journal 49(4):587–93.

⁴¹⁴ www.revolvy.com/page/George-E.-Smith

⁴¹⁵ www.megapixel.co.il/english/archive/35884

⁴¹⁶ www.fujifilm.com/innovation/achievements/ds-1p/

⁴¹⁷ Gordon E. Moore, 1965, “Cramming More Components onto Integrated Circuits,” Electronics 38(8):114–17.

⁴¹⁸ https://jpeg.org/jpeg/index.html

⁴¹⁹ www.cipa.jp/std/documents/e/DC-008-2012_E.pdf

⁴²⁰ Nils J. Nilsson, 2009, The Quest for Artificial Intelligence: A History of Ideas and Achievements, Cambridge: Cambridge University Press.

⁴²¹ www.apple.com/iphone-xr/cameras/

⁴²² ChengXiang Zhai and Sean Massung, 2016, Text Data Management and Analysis: A Practical Introduction to Information Retrieval and Text Mining, Williston, VT: ACM and Morgan & Claypool, pp. 34–36.

⁴²³ Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, 2012, “Imagenet Classification with Deep Convolutional Neural Networks,” Advances in Neural Information Processing Systems 25:1097–105.

⁴²⁴ David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams, 1986, “Learning Representations by Back-Propagating Errors,” Nature 323:533–36.

---

Kalantzis, Mary, and Bill Cope. 2020. Adding Sense: Context and Interest in a Grammar of Multimodal Meaning. Cambridge UK: Cambridge University Press. pp. 156–160. https://doi.org/10.1017/9781108862059

Digital Images and “Artificial Intelligence” §AS1.4.6c