It’s about time someone developed an anime series about deep learning. In the last several years, I’ve paid close attention to deep learning. And while I’m far from an expert on anime, I’ve watched a nonzero number of anime cartoons. And yet through neither route did I encounter even one single anime about deep learning.
There were some close calls. Ghost in the Shell gives a vague pretense of addressing AI. But the character might as well be a body-jumping alien. Nothing in this story speaks to the reality of machine learning research.
In Knights of Sidonia, if you can muster the superhuman endurance required to follow the series past its only interesting season, you’ll eventually find out that the flying space-ship made out of remnants of Earth on which Tanikaze and friends photosynthesize, while taking breaks from fighting space monsters, while wearing space-faring versions of mecha suits … [breath] contains an artificially intelligent brain-emulating parasitic nematode. But no serious consideration of ML appears.
If you were looking to anime for a critical discourse on artificial intelligence, until recently you’d be disappointed.
But that’s all changed this year with Death Note, an anime series based on a Manga by Tsugumi Ohba and available now on Netflix.
If you haven’t seen Death Note, you might not recognize its cutting critique of the deep learning craze. Metaphors for deep learning and the very core ideas and problem formulations of machine learning itself
are embedded deeply into the narrative. But strangely, this fact appears to be lost on the entire audience for this series, including the machine learning researchers among them, leaving me no choice but to provide the relevant exigesis.
Death Note is set in modern-day Japan. However, in addition to the terrestrial world, Death Note’s narrative also addresses its interaction with the “Shinigami Realm”, a wasteland populated by monstrous grim reapers who extend their own lives by taking the lives of humans.
Each Shinigami possesses a book, called a Death Note. The Shinigami can take any human’s life by writing that human’s name on paper from the Death Note. Strangely, the Shinigami aren’t quite sure why they have this power. They do not possess any special cognitive faculty lacked by ordinary humans and moreover, they appear to have no theory, from first principles that explains why the Death Note works at all.
Clearly, from the start, we are meant to understand the Death Note
as the modern deep learning toolbox. Just as the Shinigami are not quite sure why the Death Note works, so too are deep learning researchers left to speculate about the precise workings of their tools. Moreover, just as the Death Note facilitates a parasitic existence, in which the Shinigami prolong their lives by feeding on humans, deep learning methods are poised to expand the economic fortunes of a small number of practitioners, parasitically devouring millions of jobs in the process.
Transfer of the Note to the Human World
At the outset of the series, one Shinigami, Ryuk, loses his Death Note in the human world after passing through a portal to the human realm. It’s discovered by a human teenage boy named Light. It turns out that once a Shinigami loses its note in the human world, the human who finds it retains control of it. Here, Ohba takes a page from Zuckerberg. Death Notes, like information, want to be free.
Discovering the book’s magic, Light becomes intoxicated with power, using it to murder countless souls that he deems unworthy. Somehow he’s able to describe this un-ironically as making the world a better place.
To any observant reader, it’s no secret that the portal to the human realm is a thinly veiled metaphor for popular code-sharing website GitHub.com.
The loss of the Death Note symbolizes the open-sourcing of deep learning models. Light then, is a self-important, script-kiddie who happened to stumble upon an Andrej Karpathy blog post about LSTMs and now is convinced that he can (and should) replace an entire industry . All he needs to do is type “python run.py –data <filename>“. Point the weapon at the right file and he’s ready to eliminate a job. Like the dullest witted of Silicon Valley denizen, Light believes uncritically that he’s saving the world.
The Shinigami Eyes
It’s also revealed that the holder of the death note must know the name and picture the face of his/her victim. It’s not enough to just see the face.
However, the Shinigami have no such restriction. They possess “the Shinigami Eyes”, an ability that allows them to use the Death Note more efficiently. Shinigami eyes can be transferred to a human, but at the expense of half of their remaining years of life.
Death Note reminds us that it’s not enough to possess a powerful model, and to have a great dataset at hand. A practitioner must also find the right set of hyper-parameters (architecture, learning rate, choice of optimizer) to have a shot at effectively training the net. The real-life Shinigami Eyes, i.e. the black magic intuition that allows some deep learning researchers to excel at empirical work, Ohba suggests, comes at considerable expense. Perhaps he’s referring to an inability to interact socially with other humans.
Soon, Light begins using the Death Note indiscriminately, killing nearly every criminal whose name is announced on television. He fancies himself a God, and Japan’s elite police forces are forced to investigate. Given the bizarre nature of the killing, they invite pseudonymous teenage boy-wonder detective L to their side. L begins investigating the murders,
which are so plentiful that he can take a statistical approach.
He sets to work searching for patterns in the data, quickly discovering that nearly all the murders occurred after 4pm. From this, he suspects that “Kira” (as the murderer is now known, because it sounds like “killer”) might be a student. If at this point, even if you hadn’t yet realized this show’s connection to Silicon Valley, its obsession with precocious young boys might give it away.
Once it’s announced that L is investigating the Kira case, Light, being quite intelligent himself, decides to throw him off the scent. He surmises that L would notice the peculiar times of the killings, and starts scheduling the killings at random times, violating the i.i.d. assumption and dealing L a serious lesson on distributional shift.
With this stroke, the two are officially locked in an adversarial game. Note that it is not a zero-sum game as because some outcomes are strictly crap for both players. For example, it’s possible that they might both die. but along some restricted dimensions, say whether or not Light is captured, the two are indeed locked in a minimax game. Factoring in their auxiliary desiderata, the game might best be described as semi-cooperative.
I haven’t finished the series, and don’t know if I will. But I suspect the later seasons will comprise a treatise on variational methods.