Dr. Deeplearning or:
How I learned to accept failure and love feedback
Stanley Kubrick, Peter Sellers, Slim Pickens, and George C Scott have given us one of the more memorable dark comedies in history. It wasn’t supposed to be. It was probably not supposed to be an allegory for Deep Learning either, but then we can always blame that on the fluoride in the water.
General Jack D. Ripper: Fluoridation is the most monstrously conceived and dangerous communist plot we have ever had to face.
Dr Strangelove or: How I Learned to Stop Worrying and Love the Bomb premiered in 1964. It was loosely based on the book Red Alert. It is the story of a rogue general, a bumbling president, an insane major, and the mad genius Dr. Strangelove. But like most great analytics, it is not so much about the What of the movie as the How.
Anything created by Stanley Kubrick tends to be a work of love, strange love. Kubrick was so far outside the box, it is unclear if he ever noticed that one existed. As noted — this signature movie was inspired by the book Red Alert, only Red Alert did include the Dr. Strangelove character, nor was it a comedy in any sense. Kubrick, however, found it uncontrollably humorous.
Red Alert wasn’t even the original name of the book. It was Two Hours To Doom. The author was Peter George, Bryant was a pseudonym. The book and author were renamed when it was republished in America.
Kubrick, failing daily at keeping things serious, simply changed the entire concept. When the world is funny — write a comedy.
Next came Peter Sellers. Sellers had recently starred in another Kubrick film Lolita. In it, his character assumed multiple identities. Seeing no reason to change a good thing, Columbia studios insisted he play multiple parts in the new film. If at first you succeed, just keep doing it.
Unfortunately, Sellers failed at southern accents. So Slim Pickens was brought in to play Major “King” Kong. Interestingly enough, no one ever told Pickens the film was a comedy. How this fact remained hidden to him is beyond me, but by many accounts he really did little true acting anyway. He just played himself.
Now George C. Scott knew he was in a comedy, but the whole thing seemed too over-the-top to him. He attempted to reign in his character to keep things more believable. Kubrick responded by having Scott do acting exercises of his role, exaggerating his character as a “prepratory” take before a final cut. Unknown to Scott, Kubrick used all the prep takes in the film. Needless to say, this hidden agenda did not sit well with Scott, who remained angry years after the release.
Kubrick’s work was laced with hidden layers, failures, adaptations, multiple identities and individual actors often unaware of the final product (let’s call them nodes). The entire movie was an attempt to exploit and optimize maximum feedback. It was an exercise in Deep Learning, a very human one.
Dr Strangelove
As for our title character, he too had pseudonyms, hidden agendas, failures, and exaggerated feedback. As noted — he wasn’t even in the original book and yet he emerged, not just as a great character — but as the title character.
It is probably apropos that he was actually modeled after Von Neumann, a scientist at the heart of machine learning and artificial intelligence. Von Neumann also worked in game theory, work that drove him MAD. MAD being an acronym for Mutual Assured Destruction. While not the lone inspiration, remember multiple identities, Von Neumann was one of the biggest.
Von Neumann was actually known as a top-down mathematician. Strangelove (the character) being over-the-top is perhaps an homage to this fact. Deep Learning, on the other hand, is a bottoms-up process — as was Strangelove (the movie). Whatever Kubrick had intended the day he signed-on to make this movie, the outcome must have surprised him.
And so, we have the perfect analogy for Deep Learning; a bottoms-up process with nodes playing multiple identities, individual players unaware of their larger parts, optimized feedback, embraced failure, emergent outcomes, and plenty of hidden layers. I bet you weren’t expecting that.
