Prediction Machine Magic

Prediction Machine Magic What do Harry Potter, Snow White, and Macbeth have in common? These characters are all motivated by a prophecy, a prediction. Even in The Matrix, a film seemingly about intelligent machines, the human characters’ belief in predictions drives the plot. From religion to fairy tales, knowledge of the future is consequential. Predictions affect behavior. They influence decisions. The ancient Greeks revered their many oracles for an apparent ability to predict, sometimes in riddles that fooled the questioners. For example, King Croesus of Lydia was considering a risky assault on the Persian Empire. The king did not trust any particular oracle, so he decided to test each before asking for advice about attacking Persia. He sent messengers to each oracle. On the hundredth day, the messengers were to ask the various oracles what Croesus was doing at that moment. The oracle at Delphi predicted most accurately, so the king asked for and trusted its prophecy.1 As in Croesus’s case, predictions can be about the present. We predict whether a current credit card transaction is legitimate or fraudulent, whether a tumor in a medical image is malignant or benign, whether the person looking into the iPhone camera is the owner or not. Despite its Latin root verb (praedicere, meaning to make known beforehand), our cultural understanding of prediction emphasizes the ability to see otherwise hidden information, whether in the past, present, or future. The crystal ball is perhaps the most familiar symbol of magical prediction. While we may associate crystal balls with fortune-tellers predicting someone’s future wealth or love life, in The Wizard of Oz, the crystal ball allowed Dorothy to see Auntie Em in the present. This brings us to our definition of prediction: PREDICTION is the process of filling in missing information. Prediction takes information you have, often called “data,” and uses it to generate information you don’t have. The Magic of Prediction Several years ago, Avi (one of the authors) noticed a large, unusual transaction in a Las Vegas casino on his credit card. He hadn’t been in Las Vegas. He had only been there once a long time before; the losing bet of gambling doesn’t appeal to his economist way of seeing the world. After an extensive conversation, his card provider reversed the transaction and replaced the card. Recently, a similar problem occurred. Someone had used Avi’s credit card for a purchase. This time Avi didn’t see it on his statement and didn’t have to deal with the painstaking process of explaining it to a polite but firm customer service representative. Instead, he received a proactive call that his card had been compromised and that a new card was already in the mail. The credit card provider had accurately inferred, based on Avi’s spending habits and a myriad of other available data, that the transaction was fraudulent. The credit card company was so confident that they did not even block his card for a few days while they carried out an investigation. Instead, like magic, the company sent a replacement without his having to do anything. Of course, the credit card provider did not have a crystal ball. It had data and a good predictive model: a prediction machine. Better prediction allowed it to reduce fraud while, as Ajay Bhalla, Mastercard’s president of enterprise risk and security, put it, “solving a major consumer pain point of being falsely declined.”2 Business applications are well aligned with our definition of prediction as the process of filling in missing information. Credit card networks find it is useful to know whether a recent credit card transaction is fraudulent. The card network uses information about past fraudulent (and nonfraudulent) transactions to predict whether a particular recent transaction is fraudulent. If so, then the credit card provider can prevent future transactions on that card and, if the prediction is made quickly enough, then, perhaps even the current one. This notion—taking information of one kind and turning it into information of another kind—is at the heart of one of AI’s recent main achievements: language translation, a goal that has been around for all of human civilization, even enshrined in the millennia-old story of the Tower of Babel. Historically, the approach to automatic language translation was to hire a linguist—an expert on the rules of language—to exposit rules and translate them into a way they could be programmed.3 This is how, for instance, you might take a Spanish phrase and, beyond simply substituting word for word, understand that you need to swap the order of nouns and adjectives to make it a readable English sentence. The recent advances in AI, however, have enabled us to recast translation as a prediction problem. We can see the seemingly magical nature of the use of prediction for translation in the sudden change in the quality of Google’s translation service. Ernest Hemingway’s The Snows of Kilimanjaro begins beautifully: Kilimanjaro is a snow-covered mountain 19,710 feet high, and is said to be the highest mountain in Africa. One day in November 2016, in translating a Japanese version of Hemingway’s classic short story into English via Google, Professor Jun Rekimoto, a computer scientist at the University of Tokyo, read: Kilimanjaro is 19,710 feet of the mountain covered with snow, and it is said that the highest mountain in Africa. The next day, the Google translation read: Kilimanjaro is a mountain of 19,710 feet covered with snow and is said to be the highest mountain in Africa. The difference was stark. Overnight, the translation had gone from clearly automated and clunky to a coherent sentence, from someone struggling with a dictionary to seemingly fluent in both languages. Admittedly, it wasn’t quite at the Hemingway level, but the improvement was extraordinary. Babel appeared to have returned. And this change was no accident or quirk. Google had revamped the engine underlying its translation product to take advantage of the recent advances in AI that are our focus here. Specifically, Google’s translation service now relied on deep learning to supercharge prediction. Language translation from English to Japanese is about predicting the Japanese words and phrases that match the English. The missing information to be predicted is the set of Japanese words and the order in which they belong. Take data from a foreign language and predict the correct set of words in the right order in a language you know, and then you can understand another language. Do it really well, and you might not realize translation is taking place at all. Companies have wasted no time in putting this magical technology to commercial use. For example, over 500 million people in China already use a deep learning–powered service developed by iFlytek to translate, transcribe, and communicate using natural language. Landlords use it to communicate with tenants in other languages, hospital patients use it to communicate with robots for directions, doctors use it to dictate a patient’s medical details, and drivers use it to communicate with their vehicles.4 The more the AI is used, the more data it collects, the more it learns, and the better it becomes. With so many users, the AI is improving rapidly. How Much Better Is Prediction Than It Used to Be? The changes in Google Translate illustrate how machine learning (of which deep learning is a subfield) has dramatically reduced the costs of qualityadjusted prediction. For the same cost in terms of computational capacity, Google can now provide higher-quality translations. The cost of producing the same quality of prediction has dropped significantly. Innovations in prediction technology are having an impact on areas traditionally associated with forecasting, such as fraud detection. Credit card fraud detection has improved so much that credit card companies detect and address fraud before we notice anything amiss. Still, this improvement seems incremental. In the late 1990s, the leading methods