Why I’m Not Impressed With AI and Deep Learning
Something has always bugged me about the AI. I am somewhere between “mildly curious” and “skeptic” whenever I hear the future is rooted in computers like neural networks, eliminating all jobs and running us over like Terminator.
The mildly curious side of me subscribed to MIT Technology Review (the paper edition!!) and the skeptic side was pretty excited to read an article titled“Is AI Riding a One-Trick Pony?” in the Nov/Dec 2017 issue.
Before this article, the most I knew about deep learning was that you needed to feed a SHIT-TON (that’s a technical term 💩 😁) of data into a system that would work through a series of decisions to determine what is or is not a hot dog. My colleague Yi-Ying has written about the inherent biases in this training process that depends on the inputs — if all the data comes from white guys, we’re stuck only being able to serve white guys.
Reading this article, I finally learned how the heck that “training” happened. The effort behind backprop (the way programmers rejigger the connection strengths in the neutral networks game of 20,000(?) questions to avoid sharing bad answers). I have huge respect for solving that painful process of working backwards to “read” a system that’s writing itself. I learned about these “vectors” — fascinating, (almost) organic creation of representations of ideas. The word dork in me loves the connotations implied from using a mathematical concept to describe the fuzzy concept of “same”.
But, here’s the quote that really resonated with my uneasiness around the concepts:
“Deep learning in some ways mimics what goes on in the human brain, but only in a shallow way — which perhaps explains why its intelligence can sometimes seem so shallow. Indeed, backprop wasn’t discovered by probing deep into the brain, decoding thought itself; it grew out of models of how animals learn by trial and error in old classical-conditioning experiments. And most of the big leaps that came about as it developed didn’t involve some new insight about neuroscience; they were technical improvements, reached by years of mathematics and engineering.”
In my 4 years as a dog trainer at Petsmart (the coolest high school and college job ever!!), I taught pet owners how to condition their dogs to what “sit” and “stay” meant. A few dog owners were particularly concerned about the accessibility of English and insisted on teaching their dogs the commands in German — so only they could tell their dog what to do.
A fellow trainer and I joked about teaching the dogs commands with random words: “hotdog” represented “sit” and waffles would mean to “lay down”. Since I’ve yet to meet a dog born to know English, as trainers we were building these language associations. This didn’t make dogs unintelligent, it did just created a communication barrier. Fundamentally, we were teaching the dogs to relate to us using our language, and that good things came to them by paying attention. Thankfully, we used positive reinforcement (so no hitting dogs here), but the training was rooted in the ability to recognize the pattern of “I say a thing, you do a thing, you get treats”.
If you’ve ever noticed that pet owner that has to repeat the command “sit” 4 times before the dog sits, it’s not that the dog isn’t listening. The more likely culprit is the owner may have actually taught that the command pattern is “sit sit sit sit” before the dog is rewarded or forced to comply. Hand signals also play an important role; if you usually lift your hand and say “sit” at the same time, the dog may not comply if you only say “sit” without the hand gesture. In general, we noticed that dogs pick up the hand signals more quickly since it’s a communication method more in line with the dog world.
The phrase “Deep learning” to me implies a rich understanding of context and situation; but it doesn’t feel that way right now. Is it considered “deep” because we are giving a computer millions and millions of images that a human could never process? Is it “learning” because the program creates these seemingly magical categorizations on its own and references them later? But like a dog, if you change the command to something not already in the dataset, will it know how to “roll over”?
I understand how excited people get when something starts to work. It’s the same as when a dog stares lovingly into your eyes and does a perfect “sit” just to get more treats. We (designers, engineers, data scientists, Silicon Valley types) need to recognize that just because we’ve got a tool that sort of works, it still has limitations. The world’s problems are not going to be solved by the ability to train programs to “sit” and “stay” and “speak” and “find the ball”, we’ve got to think bigger than that to the commands we didn’t know we wanted.