Tag Archives: ai

Neural Streams of Consciousness

Style-extraction algorithms having reached the level of popular smartphone apps which could take the small-scale features of Hokusai’s wave or a Lichtenstein cartoon and apply them to a picture of one’s pet, it was only a matter of time before the technique was successfully applied to textual, rather than graphic, works. These first neural networks were mere mimics, more sophisticated versions of elementary Markov chains, which could produce plausible but nonsensical imitations of existing texts with no semantic content.

A breakthrough came with the Antal functor, which used a form of iterated adversarial machine learning algorithm to extrapolate multiple versions of a given text along many dimensions and then aggressively prune this ramifying cluster of words into “fixed points”, an unfortunate piece of mathematical jargon for what could be quite subtle and profound features of the source material. (The story of the functor’s use in extracting “virtual characters” from apparently objective and non-fictional texts, and the subsequent effects of this discovery on journalism and politics, have been told elsewhere.) Once this basic technique had been mastered, it could then be applied in an analogous way to that used in vision, sorting the qualities of a text on an approximate scale which ranged from such minutiae as idiosyncrasies of spelling or word frequency, to the characteristic syntactic patterns employed, and then on, with decreasing accuracy, into such large-scale qualities as extended metaphors, symbolic structures and plot.

The most famous application of this technique was the urDay service. The user registered his or her various social media accounts with urDay and allowed it to apply a battery of neural functors to the texts and images which flowed from them, taking these as a modern and technologically-mediated version of the stream of consciousness which had been pioneered in literature by Woolf and Joyce. The abstract versions of these could then be expressed in any number of ways: to generate wry or amusing animations with a cast of adorable algorithmically-generated mascots; inserted into an ever-changing roster of movie clips and viral videos as sarcastic commentary, witty cameo or heartfelt dialogues.

A set of textual plugins had been provided, more out of the curiosity of some of the development team and a sense of pride in their antecedence than any hope that urDay would have any serious impact on literary studies, much less kindle in its users a love of high modernism. With these, one could project the narration of one’s life in a kaleidoscope of styles and voices, just as Joyce had done in Ulysses: a cursory description of an annoying planning workshop or visit to a supermarket could be recounted in the language of high fantasy or science fiction. Use and abuse of these textual plugins became popular in certain literary circles, all the more because of the occasional thunderings against this digital prostitution of the art and craft of writing which came from the stodgier journals: although their output was, if anything, too facile and polished to really be groundbreaking as generated textual art, their use signalled that one was not above a certain populist bravado.

As is only natural, after a few years such collaborations seemed painfully dated, and the professional writers abandoned the field to those amateurs who enjoyed running an autoblog which gathered up and retold the output of their various encounters and days in the manner of, for example, a noir detective story, or an epic battle across frozen tundra, or a stylish psychodrama.

Their remained the matter of what became known as “the puzzles”. Certain scholars who had shifted from collaborating with the urDay plugins to analysing their outputs claimed that motifs and images seemed to be following patterns which, though elusive, could neither be attributed to the social media inputs, nor to the literary models used to generate the various styles. (The use of functorial analysis allowed this to be done with a degree of confidence.) For example, a week-long sequence from a university student’s autoblog, which alternated between a somewhat archaic translation of Sei Shonagon’s Pillow Book and Patti Smith’s memoirs, showed a striking affinity with certain of Pound’s Pisan Cantos, a work which neither the student nor her chosen electronic amanuenses had any connection. An archaeologist’s field notes, transformed into an elaborate science-fantasy scenario, spontaneously revealed a correspondence between certain ruins on the shores of the Persian Gulf and the galactic coordinates of active pulsars. Once one began looking for such patterns, it was said, they began to emerge everywhere, and perhaps it was this sense of ubiquity which explained the somewhat tepid response with which these demonstrations were greeted. While happy enough, at least in some circles, to let the false leads and teasing traps laid down by a legitimate genius like Joyce keep them busy for centuries to come, literary scholars saw the apparently limitless sea of neural “puzzles” as nothing more than an epiphenomenon of their computational origin, as uninteresting to them as the technical details of the programming languages used to create their word processors or functorial analysers.

Eventually, the “puzzles” became the hunting ground of that even more prolific realm of amateurism, the conspiracy theorists, to be added to their never-ending roster of patterns and coincidences, world without end.

Deepdreams Dive Two: the MIT Places Neural Net

Here are some more sample images run on the same randomised coloured fog as the last dive. These are done with a different neural net, based on the MIT Places database. They lack the biomorphic horrors of the default net, but they have a kind of weird beauty.

3a_512

3a. Low-level features, and a tendency to diffract colour

3b_512

3b. These look curiously suggestive of the spiral patterns on ancient Celtic artefacts.

4a_512

4a. You can really see how the Places net gets cues from colour here: as you might expect, the green parts seem to want to turn into parks.

4b_512

4b. This layer is just crazy for chairs. And windfarms.

4c_512

4c. Pagodas as far as the eye can see.

4d_512

4d. The wind farms are back, and the pagodas are starting to look like stupas.

4e_512

4e. It’s strange how at the higher levels, this model starts to go wonky. There are anphitheatres forming in the lower left.

5a_512

5a. Oddly specific details here: a watertower and a fountain.

5b_512

5b. This is the only layer in this net that I find very unsettling. There are suggestions of distorted faces, and something like a pine-forest in the middle.