“I see your infinite form in every direction,
with countless arms, stomachs, faces, and eyes.”
~ Bhagavad-Gītā 11 16
What the hell is that thing? Our sensemaking gear immediately kicks into overdrive. If Cthulhu had had a pet slug, this might be what it looked like. But as you look deeper into the picture, all sorts of other things begin to emerge. In the lower left-hand corner there are buildings and people, and people sitting on buildings which might themselves be on wheels. The bottom center of the picture seems to be occupied by some sort of a lurid, lime-colored fish. In the upper right-hand corner, half-formed faces peer out of chalices. The background wallpaper evokes an unholy copulation of brain coral and astrakhan fur. And still there are more faces, or at least eyes. There are indeed more eyes than an Alex Grey painting, and they hew to none of the neat symmetries that make for a safe world. In fact, the deeper you go into the picture, the less perspective seems to matter, as solid surfaces dissolve into further cascades of phantasmagoria. The same effect applies to the principal thing, which has not just an indeterminate number of eyes, ears or noses, but even heads.
The title of the thread wasn't very helpful, either: “This image was generated by a computer on its own (from a friend working on AI)”. For a few days, that was all anyone knew, but it was enough to incite another minor-scale freakout about the nature and impending arrival of Our Computer Overlords. Just as we are helpless to not over-interpret the initial picture, so we are all too willing to titillate ourselves with alarmist speculations concerning its provenance. This was presented as a glimpse into the psychedelic abyss of artificial intelligence; an unspeakable, inscrutable intellect briefly showed us its cards, and it was disquieting, to put it mildly. Is that what AI thinks life looks like? Or stated even more anxiously, is that what AI thinks life should look like?
*
Alas, our giddy Lovecraftian fantasies weren't allowed to run amok for more than a few days, since the boffins at Google tipped their hand with a blog post describing what was going on. The image, along with many others, were the result of a few engineers playing around with neural networks, and seeing how far they could push them. In this case, a neural network is ‘trained' to recognize something when it is fed thousands of instances of that thing. So if the engineers want to train a neural network to recognize the image of a dog, they will keep feeding it images of the same, until it acquires the ability to identify dogs in pictures it hasn't seen before. For the purposes of this essay, I'll just leave it at that, but here is a good explanation of how neural networks ‘learn'.
The networks in question were trained to recognize animals, people and architecture. But things got interesting when the Google engineers took a trained neural net and fed it only one input – over and over again. Once slightly modified, the image was then re-submitted to the network. If it were possible to imagine the network having a conversation with itself, it may go something like this:
First pass: Ok, I'm pretty good at finding squirrels and dogs and fish. Does this picture have any of these things in it? Hmmm, no, although that little blob looks like it might be the eye of one of those animals. I'll make a note of that. Also that lighter bit looks like fur. Yeah. Fur.
Second pass: Hey, that blob definitely looks like an eye. I'll sharpen it up so that it's more eye-like, since that's obviously what it is. Also, that fur could look furrier.
Third pass: That eye looks like it might go with that other eye that's not that far off. That other dark bit in between might just be the nose that I'd need to make it a dog. Oh wow – it is a dog! Amazing.
The results are essentially thousands of such decisions made across dozens of layers of the network. Each layer of ‘neurons' hands over its interpretation to the next layer up the hierarchy, and a final decision of what to emphasize or de-emphasize is made by the last layer. The fact that half of a squirrel's face may be interpellated within the features of the dog's face is, in the end, irrelevant.
But I also feel very wary about having written this fantasy monologue, since framing the computational process as a narrative is something that makes sense to us, but in fact isn't necessarily true. By way of comparison, the philosopher Jacques Derrida was insanely careful about stating what he could claim in any given act of writing, and did so while he was writing. Much to the consternation of many of his readers, this act of deconstructing the text as he was writing it was nevertheless required for him to be accurate in making his claims. Similarly, while the anthropomorphic cheat is perhaps the most direct way of illustrating how AI ‘works', it is also very seductive and misleading. I offer up the above with the exhortation that there is no thinking going on. There is no goofy conversation. There is iteration, and interpretation, and ultimately but entirely tangentially, weirdness. The neural network doesn't think it's weird, however. The neural network doesn't think anything, at least not in the overly generous way in which we deploy that word.
I believe that there are more accurate metaphors to be had, ones that rely on narrower views of our subjectivity, not the AI's. For example, there is the children's game of telephone, where a phrase is whispered from one ear to the next. Given enough iterations, what emerges is a garbled, nonsensical mangling of the original, but one that is hopefully still entertaining. But if it amuses, this is precisely because it remains within the realm of language. The last person does not recite a random string of alphanumeric characters. Rather, our drive to recognize patterns, also known as apophenia, yields something that can still be spoken. It is just weird enough, which is a fine balance indeed.
The world of sound also provides a metaphor emphasizing our tendency towards apophenia. Take a moment to listen to this clip. As the accompanying BBC article elaborates:
What did you hear? To me, it sounds obvious that a female voice is repeating “no way” to oblivion. But other listeners have variously reported window, welcome, love me, run away, no brain, rainbow, raincoat, bueno, nombre, when oh when, mango, window pane, Broadway, Reno, melting, or Rogaine.
This illustrates the way that our expectations shape our perception…. We are expecting to hear words, and so our mind morphs the ambiguous input into something more recognisable. The power of expectation might also underlie those embarrassing situations where you mishear a mumbled comment, or even explain the spirit voices that sometimes leap out of the static on ghost hunting programmes.
Even more radical are Steve Reich's tape loop pieces, which explore the effects of when a sound gradually goes out of phase with itself. In fact, 2016 will be the 50th anniversary of “Come Out“, one of the seminal explorations of this idea. While the initial phrase is easy to understand, as the gap in phase widens we struggle to maintain its legibility. Not long into the piece, the words are effectively erased, and we find ourselves swimming in waves of pure sound. Nevertheless, our mental apparatus stills seeks to make some sort of sense of it all, it's just that the patterns don't obtain for long enough in order for a specific interpretation to persist.
Of course, the list of contraptions meant to isolate and provoke our apophenic tendencies is substantial, and oftentimes touted as having therapeutic benefits. We slide into sensory deprivation tanks to gape at the universe within, and assemble mail-order DIY ‘brain machines' to ‘expand our brain's technical skills'. This is mostly bunk, but all are predicated on the idea that the brain will produce its own stimuli when external ones are absent, or if there is only a narrow band of stimulus available. In the end, what we experience here is not so much an epiphany, as apophany.
*
In effect, what Google's engineers have fabricated is an apophenic doomsday machine. It does one thing – search for patterns in the ways in which it knows how – and it does those things very, very well. A neural network trained to identify animals will not suddenly begin to find architectural features in a given input image. It will, if given the picture of a building façade, find all sorts of animals that, in its judgment, already lurk there. The networks are even capable of teasing out the images with which they are familiar if given a completely random picture – the graphic equivalent of static. These are perhaps the most compelling images of all. It's the equivalent of putting a neural network in an isolation tank. But is it? The slide into anthropomorphism is so effortless.
And although the Google blog post isn't clear on this, I suspect that there is also no clear point at which the network is ‘finished'. An intrinsic part of thinking is knowing when to stop, whereas iteration needs some sort of condition wrapped around the loop, otherwise it will never end. You don't tell a computer to just keep adding numbers, you tell it to add only the first 100 numbers you give it. Otherwise the damned thing won't stop. The engineers ran the iterations up until a certain point, and it doesn't really matter if that point was determined by a pre-existing test condition (eg, ‘10,000 iterations') or a snap aesthetic judgment (eg, ‘This is maximum weirdness!'). The fact is that human judgment is the wrapper around the process that creates these images. So if we consider that a fundamental feature of thinking is knowing when to stop doing so, then we find this trait lacking in this particular application of neural networks.
In addition to knowing when to stop, there is another critical aspect of thinking as we know it, and that is forgetting. In ‘Funes el memorioso', Jorge Luis Borges speculated on the crippling consequences of a memory so perfect that nothing was ever lost. Among other things, the protagonist Funes can only live a life immersed in an ocean of detail, “incapable of general, platonic ideas”. In order to make patterns, we have to privilege one thing over another, and dismiss vast quantities of sensory information as irrelevant, if not outright distracting or even harmful.
When the brain is rewarded unexpectedly, dopamine surges, prompting the limbic “reward system” to take note in order to remember how to replicate the positive experience. In contrast, negative encounters deplete dopamine as a signal to avoid repeating them. This is a key learning mechanism which is also involves memory-formation and motivation. Scientists believe the brain establishes a new temporary neural network to process new stimuli. Each repetition of the same experience triggers the identical neural firing sequence along an identical neural journey, with every duplication strengthening the synaptic links among the neurons involved. Neuroscientists say, “Neurons that fire together wire together.” If this occurs enough times, a secure neural network is established, as if imprinted, and the brain can reliably access the information over time.
The hyperlearning hypothesis posits that schizophrenics have too much dopamine in their brains, too much of the time. Take the process described above and multiply it by orders of magnitude. The result is a world that a schizophrenic cannot make sense of, because literally everything is important, or no one thing is less important than anything else. There is literally no end to thinking, no conditional wrapper to bring anything to a conclusion.
Unsurprisingly, the artificial neural networks discussed above are modeled on precisely this process of reinforcement, except that the dopamine is replaced by an algorithmic stand-in. In 2011, Uli Grasemann and Risto Miikkulainen did the logical next step: they took a neural network called DISCERN and cranked up its virtual dopamine.
Grasemann and Miikkulainen began by teaching a series of simple stories to DISCERN. The stories were assimilated into DISCERN's memory in much the way the human brain stores information – not as distinct units, but as statistical relationships of words, sentences, scripts and stories.
In order to model hyperlearning, Grasemann and Miikkulainen ran the system through its paces again, but with one key parameter altered. They simulated an excessive release of dopamine by increasing the system's learning rate — essentially telling it to stop forgetting so much.
After being re-trained with the elevated learning rate, DISCERN began putting itself at the center of fantastical, delusional stories that incorporated elements from other stories it had been told to recall. In one answer, for instance, DISCERN claimed responsibility for a terrorist bombing.
Even though I find this infinitely more terrifying than a neural net's ability to create a picture of a multi-headed dog-slug-squirrel, I still contend that there is no thinking going on, as we would like to imagine it. And we would very much like to imagine it: even the article cited above has as its headline ‘Scientists Afflict Computers with Schizophrenia to Better Understand the Human Brain'. It's almost as if schizophrenia is something you can pack into a syringe, virtual or otherwise, and inject it into the neural network of your choice, virtual or otherwise. (The actual peer-reviewed article is more soberly titled ‘Using computational patients to evaluate illness mechanisms in schizophrenia'.) We would be much better off understanding these neural networks as tools that provide us with a snapshot of a particular and narrow process. They are no more anthropomorphic than the shapes that clouds may suggest to us on a summer's afternoon. But we seem incapable of forgetting this. If we cannot learn to restrain our relentless pattern-seeking, consider what awaits us on the other end of the spectrum: it is not coincidental that the term ‘apophenia' was coined in 1958 by Klaus Conrad in a monograph on the inception of schizophrenia.