Superheroes of Deep Learning Vol 1: Machine Learning Yearning

Full PDFs free on GitHub. To support us, visit Patreon.

David silver, Andrew ng and Fei Fei li in their superhero form as Q-Silver, MOOC and Benchmark, respectively. Q-Silver is in the middle and is lunging towards the screen. MOOC is to the left and is jumping up into the screen with his arms outstretched and muscles in full display. Benchmark is lunging in cat-like positive to the right. Machine Learning Yearning is written above them.
Portrait of Juergen Schmidthuber in superhero form as 'The Enforcer' breaking 4th wall and pointing to the screen with a stern look on his face Terms of Use All the panels in this comic book are licensed CC BY-NC-ND 4.0. Please refer to the license page for details on how you can use this artwork. TL;DR: Feel free to use any of the artwork in your presentations/articles as long as you provide the proper citation and do not make modifications to the art. Cite as: Falaah Arif Khan and Zachary C. Lipton. “Superheroes of Deep Learning Vol 1: Machine Learning Yearning” (2020)
It's a lazy sunday on Earth. Mom and Dad sit on the kitchen table, coffee in hand. Dad is reading the 'Technology Times', while Mom keeps an eye on the kid playing in the yard outside. The kid visible through the window and is seen to be playing on a swing, while Grumpy the cat, is looking at a squirrel that is running up a tree. The 'Technology Times' in Dad's hand features a spread of the superheroes of Deep learning on the cover. On the back cover, visible to the audience, are the following advertisements (from left to right); 'Come work with the Enforcer at RNNAISNSE' with a picture of the Enforcer with his fist extending towards the screen. The Rigor Police, weeknights @8 only on Showtime and Keeping Up with the ML Kardashians [Cat stuck in a tree] Child screams "Oh Noooo" Mom throws hands up in exasperations while Dad clutches his face in worry. The child is seen crying in the background Mom says "OH MY! GRUMPY IS STUCK IN THE TREE! WHAT EVER SHOULD WE DO? "
[Dad — self-serious, scratching chin ponderously staring at headline] Well… If you had asked me at any other point in history, I would have said “call the fire department!” But times are changing… and I’ve been doing some reading. [Dad — pointing to newspaper] Look here, this article says that deep learning AIs have surpassed human beings at a wide range of problems requiring dexterous manipulation. Dad points to a newspaper with the headline 'AI Robot Hand: Today Rubik's Cube, Tomorrow the Real World' Dad says 'This just might be a job for the Deep Learning Superheroes!' and looks up with a resolute look on his face, while Mom and Kid look excited
[MOOC arrives on the scene, to the rescue. Student army follows in his wake, so many students, they are all anonymous and trailing off far into the distance. Few-to-none of them are individually recognizable, just a mass of people] [MOOC]: Hello there, humans. I heard you were in some kind of trouble. Some of my students and I thought we’d stop by to lend a hand. [MOOC gestures at infinitude of bodies stretching into the sunset]
We see grumpy unhappily sitting in the tree. Mom and kid and MOOC look up at her from below. MOOC says "Please articulate the nature of your task?" Mom, gestures towards the tree and says "Our beloved Grumpy chased a squirrel up the oak tree, and can’t get down by herself!" MOOC—turns to class and says: "Remember kids, most humans aren’t fluent in the ML lingo. You need to ask the right questions to get to the heart of the problem." MOOC, excitedly throws his hands into the air and screams at Mom and Dad:“WHAT ARE YOUR INPUTS? WHAT ARE YOUR OUTPUTS? WHAT IS THE LOSS FUNCTION? STATE YOUR INDUCTIVE BIASES! GIVE US YOUR DATASET” Family — looks to each other, perplexed and say, "… dataset?"
MOOC, doing the 'Think-flex', says, "The situation is even worse than I thought. Looks like we’re going to have to formulate the problem ourselves." Behind him is an army of students, clawing at their eyes MOOC, decidedly, "OK, the first thing we’re going to need is a dataset." Sky darkens, lightning bolt strikes in distance, and makes a 'CRACK' sound
Benchmark descends from the sky, arms extended as she controls data and says 'Did somebody say...Dataset?'
Benchmark summons data into the shape of a workstation, with 3 monitors. She excitedly starts typing on the keyboard and is glued to the screen as she says, "Fortunately, all the data you need is already on the internet, if only you know where to look". [emphatically hits ENTER] "There! I’ve pulled every satellite photograph of this tree since the covert beginnings of orbital surveillance during the Kennedy administration." MOOC smiles as he looks at the all data Benchmark has pulled.
Benchmark, looking at MOOC, says "Now we just need to hire an army of crowdworkers to label the images as “cat / no cat”." A mass of students stand in the distance, drops of sweat forming as they realize they’ve been conscripted into hours of Turk work Montage of pictures of students being presented an image and asked to answer whether there is a cat in the picture. Students look steadily more tired and the sky steadily darkens
Benchmark levitates herself into the air and says, "There you go, one million(!) images of this very tree, with ground truth cat labels for each." MOOC jumps into the air, pen drive in hand, as Benchmark summons all the labelled data into the device.
MOOC connects the pen drive to a computer and emphatically says, "Ok students, now's your time to shine". MOOC, mic-dropping the pen drive, says, "DROP THAT LOSS"
A group of students sit, with laptops, some typing excitedly, others lost in thought. The same code snipped is shown to be written on their laptops: from pytorch import favorite_cnn x_train = … y_train =   for e in epoch: for batch in load(x_train, ytrain): favorite_cnn.drop_that_loss(batch)
MOOC, holding up the arm of a student, says "And the winner—with 94.7% accuracy— is TheAlchemist! OK, let’s see what this technology can do." The student walks over to the tree, pulls out his phone, takes a picture of the tree, and says "OK, let's see what this technology can do!" Image shows up on the monitor and TheAlchemist types commands excitedly.
TheAlchemist points to the monitor and says 'Sir, ma’am, there’s definitely a cat in your tree'. MOOC looks excitedly at Mom and Dad. Dad is smiling, while Mom has her hands up in exasperation Mom, looking annoyed, levels with Dad, "It’s getting late, honey. Do you think maybe we should call the fire department?". Dad, shrugging, "The article said these guys are the best. They know what they’re doing. We just need to be patient" Behind them we see MOOC and TheAlchemist being celebrated by the other students
MOOC says to his students, "So class, that was a test and I hope you all learned a valuable lesson. Machine learning isn’t just about making predictions. Often, it’s about taking actions! " One student raises her hand and asks, "Are you talking about deep reinforcement learning?!?" Approaching from the easy is... Q-Silver Q-Silver is seen running towards the screen, surrounded by flying drones.
Q-Silver strikes a solemn superhero pose and says, "That’s right, kids! With the combined powers of dynamic programming and function approximation, we are going to save Grumpy!" He then throws his hand in the air, while summoning the drones and says, "I’ve brought along two quadcopters per team. Let’s get to work!"
Montage of students trying out different implementations: [Student 1] I’m going to use Q-learning! [Student 2] I’m going to use Policy Gradient! [Student 3] A3C for the win! [Student 4] Nobody has a chance against my Rainbow implementation! [Student N] Wait… what’s the reward function?! [Student N+1 — Pointing at StackOverflow] [headline: What’s the best Reward Function for Reinforcement Learning?] I've got it!
Students, hunched over their laptops: 1 Million points for rescuing the cat! 1 point for every second in the air Negative 10,0000 points for crashing Three floating heads, bearing resemblance to Benchmark, Q-Silver and MOOC appear in front of a group of students. The one in the middle says, "Alright team, show me what you got!"
[All drones lift off and fly in random directions off into the sunset] Students stand by looking and pondering, "Where are they all going?!?...Maybe they're exploring" Deep learning heroes scratch heads in puzzlement. Mom, annoyed, "Still sure about these guys?". Dad, calms her down and says, "Patience, dear"
[Q-Silver — scratching chin, thoughtful expression on face] When we built the first version of AlphaGo, we benefited from training predictive models to guess the next move based on millions of professional human games. [Student] I’ve heard of this … Imitation Learning, right? [Q-Silver] That’s right, kids. Why learn from scratch when we can kick-start our models with trajectories sampled from expert demonstrations? [Student—pensive] But where are we going to get the data? [Benchmark, assembling the students like a coach at half time] Time is of the essence. Here’s the plan. Each team must find another cat stuck in a tree and rescue it manually by piloting your drones. Don’t forget to record the complete sequence of video observations and actions taken. Let’s go!
Students scatter in all directions. Tons of drones fly around the yard. Mom and dad look absolutely befuddled. Mom asks, "If you’re capable of saving cats by manually piloting the quadcopters, why don’t you save Grumpy now?" MOOC and Q-Silver give each other a knowing look. MOOC laughs as he gestures towards Mom and Dad and says, "Humans..."
[Student, deep in thought] There don’t seem to be any other cats in trees… how do we gather training data? [A different student] I know! Let’s use the quadcopters to first put the cats IN the trees! [Montage of drones dropping cats into trees]
Student looks at the audience with a slight smile, while holding a remote control and says, "Ok… now that we have cats in trees, let’s get rescuing!" [Montage of drones rescuing cats from trees]
5 hours later... Mom, Dad and a bunch of students stand looking worriedly at Grumpy in the tree. It is dusk and Grumpy is asleep in the tree. MOOC and Q-Silver smile excitedly at the screen. MOOC, gestures a double thumbs-up at a hovering drone, while Q-Silver says. "Ok, this imitation policy should do the trick!" Q-Silver releases drone with grand gesture, like Noah releasing a dove. Drone crashes into the house and dies. The sound wakes Grumpy up.
[MOOC — knowingly] Anyone dare to guess what went wrong here? [Student, deep in thought] All of our training data was collected during the day… but we deployed the model at night? Doesn’t that violate the i.i.d. assumption? [MOOC] That’s right! [Child starts to cry] [Mom] That's it, I'm calling the fire department now [Dad, looking at the newspaper in his hand and looking utterly defeated] I just don’t see how it’s possible that the news could have overstated the capabilities of today’s AI… [MOOC] You see students, Machine Learning isn’t about replacing humans, it’s about complementing their abilities! Let’s demonstrate what humans and AI can achieve together!
The fire department finally arrives… Montage of firemen climbing up tree, while swathing away the rogue Quadcopters and holding back overly eager MOOC and Q-Silver. They finally get Grumpy down.
Fireman hands Grumpy to the family. Mom says. "Thank you so much officer. You saved Grumpy!!! We are so grateful!" Firefighter says, "Just glad we could help!" The DL Heroes jump in, unwelcome. Benchmark has summoned celebratory wine from data, Q-Silver is striking a victory pose and MOOC is standing in quintessential superhero pose (with his chest out, arms on waist) MOOC says, "You’re most welcome! Positively thrilled that we could be of assistance!". Dad smiles uneasily, where as Mom outright frowns.
Later... [Lex Fridman] The following is a conversation with MOOC, Benchmark, and Q-Silver, famed superheroes of deep learning, computer vision, and reinforcement learning. A quick summary of the ads. This episode is brought to you by HardBank, StashApp, and Ahoylent. If you enjoy this thing, subscribe on YouTube, review it with 5 stars on Apple Podcasts, follow on Spotify, support it on Patreon, or connect with me on Twitter (If you can) MOOC, Q-Silver and Benchmark are being interviewed by Lex. MOOC and Q-Silver smile, while Benchmark looks bored. Lex gushes, "Wow. Just wow. I have been such a huge fan for years. So many amazing ML techniques and algorithms came together in just the right way today in what I really think will be remembered as a defining milestone of the AGI age. " Lex is now interviewing Grumpy. Lex, asks seriously, "So, Grumpy, excuse my romanticized question, I’m Russian. I have to ask- What do you think is the meaning of life?" Grumpy growls into the microphone.
Even later... [Cut to Siraj’s new video] [Siraj Raval] Yo Grumpy — You ain’t Lumpy Stuck in that tree, nowhere to pee You lit it up girl, Burning Yearning for machine learning Eyes turning, to the sky, for drones This cat moans for the AI revolution The solution, my absolution, is convolution My algorithms fly, my rhythm so sly Don’t mean to rub it in yo faces But my complicated Hilbert spaces Are for the ages. They’re the rages. Start taking notes, you’ll need pages. [turns stares at the camera] Hello world, It's me. Word2Vec, Input in, Dot Product, Activate, Do it 1nce, do it 2wice, input out, errors done I don’t need a label, I just learnt to do it without 1 “Solve AI or die tryin” [drops mic]
A white door appears... It has the words "Quantum door" written on it. Siraj opens the door and exits through it.

Related Posts

Authors

Author: Falaah Arif Khan

An Engineer/Scientist by training and an Artist by nature, Falaah is a Research Fellow at the CVIT Lab at IIIT-Hyderabad and an Artist-in-Residence at the Montreal AI Ethics Institute.

7 thoughts on “Superheroes of Deep Learning Vol 1: Machine Learning Yearning”

Leave a Reply

Your email address will not be published. Required fields are marked *