In the second part of our “A Mathless Guide to Neural Networks”, we’ ll take a look at why high-quality, labeled data is so important, where it comes from, ..
We’ ve already taken a look at neural networks and deep learning techniques in a previous post, so now it’s time to address another major component of deep learning: data—meaning the images, videos, emails, driving patterns, phrases, objects, and so on that are used to train neural networks.
Surprisingly, despite our world being quite literally deluged by data—currently about 2.5 quintillion bytes a day, for those keeping tabs—a good chunk of it is not labeled or structured, meaning that for most current forms of supervised learning it’s unusable. And deep learning in particular depends on a steady supply of the good, structured and labeled stuff.
In the second part of our “A Mathless Guide to Neural Networks”, we’ ll take a look at why high-quality, labeled data is so important, where it comes from, how it’s used, and what solutions our eager-to-learn machines can expect in the near-term future.
Supervised learning: I wanna hold your hand
In our post about neural networks, we explained how data is fed to machines through an elaborate sausage press that dissects, analyzes, and even refines itself on the fly. This process is considered supervised learning in that the giant piles of data fed to the machines have been painstakingly labeled in advance. For example, to train a neural network to identify pictures of apples or oranges, it needs to be fed images that are labeled as such. The idea is that machines can be groomed to understand data by finding what all pictures labeled apple or orange, respectively, have in common, so they can eventually use those recognized patterns to more accurately predict what they are seeing in new images. The more labeled pictures they see, the bigger (and more diverse) the dataset, the better they can refine the accuracy of their predictions; practice makes (almost) perfect.
This approach is useful in teaching machines about visual data, and how to identify anything from photographs and video to graphics and handwriting. The obvious upside is that it is now relatively commonplace for machines to be equal or even better than humans at say, image recognition for a number of applications. For instance, Facebook’s Deep Learning software is able to match two images of an unfamiliar person at the same level of accuracy as a human (better than 97% of the time) , and Google, earlier this year, unveiled a neural network that can spot cancerous tumors in medical images more accurately than pathologists.
Unsupervised learning: Go west, young man
The companion to supervised learning, as you might guess, is called unsupervised learning. The idea is that you loosen the leash on your machine and let it dive into the data to discover and experience it on its own, look for patterns and connections, and come to conclusions, without requiring the guidance of a chaperone.
This technique had long been frowned upon by a certain segment of artificial intelligence (AI) scientists, but in 2012, Google demonstrated a deep learning network that was able to decipher cats, faces, and other objects from a giant pile of unlabeled images. This technique is impressive and produces some extremely interesting and useful results, but, so far, unsupervised learning doesn’ t match the accuracy and effectiveness of supervised training for many purposes—more on that in a bit.
Data, data, everywhere
It is in the chasm between these two techniques that we run into the larger issues that are proving to be confounding. It’s useful to liken these machines to human babies. We know that by simply setting our baby loose, without guidance, it’ ll learn, but not necessarily what we want it to learn, nor in any predictable way. But since we also teach our baby by instructing it, then we need to expose it to large numbers of objects and concepts in an essentially infinite number of topics.
We need to teach our baby about directions, animals and plants, gravity and other physical properties, reading and language, food types, and the elements, you know—the very stuff of existence. All of this can more or less be explained over time with a mix of show-and-tell and answering the endless questions that any curious young human asks.
It’s a tremendous undertaking, but one that most parents, as well as other people around the average child, take on each and every day on the fly. A neural network has the same needs, but its focus is usually more narrow and we don’ t really socialize with it, so the labels need to be much more precise.
Currently there are a number of ways that AI researchers and scientists can get access to data to train their machines. The first way is to go out there and amass a giant stockpile of labeled data on your own. This happens to be the case for companies like Google, Amazon, Baidu, Apple, Microsoft, and Facebook, all of whom have businesses that, funnily enough, generate breathtaking amounts of data—much of it laboriously curated for free by customers.
It would be folly to try and list them all here, but think of the billions of labeled and tagged images uploaded to the cloud storage of all these companies’ databases. Then think about all the documents, the search queries—by voice, and text, and photos and optical character recognition—the location data and mapping, the ratings and likes and shares, the purchases, the delivery addresses, the phone numbers and contact info and address books, and the social connections.
Legacy companies—and any company of huge scale—tend to have a unique advantage in machine learning in that they have copious amounts of specific types of data (which may or may not be valuable in the end, but often are) .
Data the hard way
If you don’ t happen to own a Fortune 100 company with collections of trillions of data points, then you’ d better be good at sharing (or have deep pockets) . Access to lots of extremely varied data is a key part of AI research. Fortunately, there is already a large number of free and publicly shared labeled datasets that cover a mind-boggling array of categories (this Wikipedia page hosts links to dozens and dozens) .
Depending on your fancy, there are datasets showing everything from human facial expressions and sign language to the faces of public figures and skin pigmentation. You can find millions of images of crowds, forests, and pets—all kinds of pets—or sift through boatloads of user and customer reviews. There are also datasets consisting of spam emails, tweets in multiple languages, blog posts, and legal case reports.
New kinds of data are emerging from the myriad of increasingly ubiquitous sensors in the world, such as medical sensors, motion sensors, smart device gyroscopes, heat sensors, and more. And then there are all those pictures people take of their food, wine labels, and ironic signage. In other words, there’s no shortage whatsoever of data in its purest form.
So what’s the problem!?
Despite this apparent cornucopia of data, in practice it turns out that many of these collections aren’ t so broadly useful. Either they are too small of a collection, they are poorly or partially labeled, or they just don’ t meet your needs. For instance, if you’ re hoping to teach a machine to recognize a Starbucks logo in images, you may only be able to find a training database of images that have been variously labeled “beverages” or “drinks” or “coffee” or “container” or “Joe.” Without the right labels, they just aren’ t so useful. And the average law firm or established corporation may have millions of millions of contracts or other paperwork in its databases, but that data isn’ t usable since it’s likely in a simple unlabeled PDF format.
Another challenge in terms of optimal data is making sure that the training sets used are both large and diverse. Why? Let’s explore the idea of training data with a simple thought experiment. Imagine we give a little kid, we’ ll call him Ned, the task of recognizing Spanish words on flashcards. When shown a flashcard, all Ned needs to do is say “Yes, this is Spanish” or “No, this is not Spanish.”
Having never seen nor spoken Spanish before, this kid Ned is given ten random flashcards in order to learn what Spanish words do and do not look like.
