“Give a man a fish and you feed him for a day; teach a man to fish and you feed him for a lifetime” reads the old proverb. After jumping on the Deep Learning bandwagon (but this applies to the Big Data, and any other bandwagon), I (my Neural Nets) have been constantly fed, and it has been great. I have been specially enchanted by how Convolutional Neural Networks (CNNs, ConvNets) have performed at classifying handwritten numbers, detecting cats, and spotting facial landmarks. The problem is that, that feast was already cooked and readily served on my (NNs) table. It was not my show, it was someone else’s.
If you are an expert in Machine Learning you will read this and think “he finally got it”. If you are just beginning and still in the tutorials (there are many, for many frameworks/packages/libraries, all convincing you they are the best) then you probably still don’t get my feeling. You might say “but the X tutorial on Y package explained very well (and pointed to nice materials) how a CNN works” and you are right, but how about the datasets? Data are the fish for our CNNs (and our RNNs and any other NN), but the data was not just fished (or hunted) for us (and our NNs), it was already nicely cooked and seasoned (cleaned, labeled, formatted).
When people claim Deep Learning will bring Machine Learning to the masses they are right, but except for a few experts, and others sitting on large amounts of data, the masses will be largely playing with MNIST, CIFAR-10/100, and other datasets available through Kaggle competitions. You still don’t know what I mean? I invite you to follow a TensorFlow tutorial (the same applies for any other library, they all use the same datasets), skip all the explanations, just copy and paste the code, then see the magic happening. After the ecstasy wears off, define a new task, not recognizing numbers or airplanes, but maybe detecting emotions in faces (like I did) and good luck my friend.
Deep Learning, like Big Data before it, is not bringing Machine Learning (or Data Science) to the masses in the right way. It feels more like politicians before the elections, coming to the less fortunate with loads of food in order to win votes (this happens all the time in my country Honduras). To really bring Machine Learning to the masses it is necessary to teach us the nice tricks that top notch scientists use to collect and prepare data. To my head come things I have done myself before like crawling data sources (Twitter, Instagram, Google, …), crowdsourcing our labeling and cleaning tasks, and other more advanced techniques. To bring Machine Learning to the masses stop feeding the masses, teach them how to fish.
This is not a critique to Deep Learning. I understand there is a large number of people with some sort of experience dealing with data, and Deep Learning can become a very useful tool for them. The raison d’être of this article is that many are irresponsibly saying Deep Learning will bring Machine Learning to the masses, and the masses will come, and the masses will get disappointed, and suffer.