LeadershipCryptocurrencyInflationGreat ResignationInvesting

How to make A.I. smarter

September 21, 2020, 1:30 PM UTC

The electronic spreadsheet has been around for about 50 years. An ingenious invention originally meant to digitize bookkeeping, the software has enabled researchers and businesspeople to input infinite rows and columns of disparate data and then analyze the information with the aid of a computer. It is such standard fare today that schoolchildren are as likely to use free spreadsheet programs as financial analysts are to manage budgets. 

What spreadsheets cannot do is think. That’s the preserve of newer, more powerful types of software called neural networks, complex artificial intelligence programs designed to mimic the computational processes of the human brain. And for reasons unique to the development of neural networks in recent years, images—rather than so-called structured data, columns and rows of text and numbers, for example—have been the preoccupation of top A.I. researchers. In other words, powerful computers can sift through millions of photos of cats to understand minute feline characteristics. But the same software struggles to intuit fields in a humble spreadsheet.

This has been deeply frustrating to data scientists in fields like medical research, finance, and operations, where structured data is the coin of the realm. The problem, researchers say, is one of emphasis as well as capabilities. “Most of data we deal with is structured, or we have imposed some kind of structure on it,” says Bayan Bruss, an applied machine learning researcher at the financial firm Capital One. “There’s this big gap between the advances in deep learning and the data that we have. A lot of what we do is try to close that gap.”

Illustration by Lena Vargas

Fledgling projects at a handful of companies are trying to bridge the divide. At biotech powerhouse Genentech, for example, data scientists recently spent months building a spreadsheet with the health records and genomic data of 55,000 cancer patients. The fields contain nuggets such as age, cholesterol levels, and heart rates, as well as more sophisticated attributes like molecular profiles and genetic abnormalities. Genentech’s plan is to feed this information into a neural network that can map a patient’s health attributes. The hoped-for outcome is a breakthrough drug that is potentially unique to each patient.

The problem is that researchers are just now beginning to teach neural networks how to consume structured data like the spreadsheets Genentech is building. “The majority of our data is structured data, whether it’s from clinical trials or electronic health records,” says Ryan Copping, global head of analytics for personalized health care data science at Genentech. If computer networks can analyze and make their own realizations about similarities among patient profiles, he says, “then you could start looking at outcomes and thinking about which patients we can target with which therapies. That’s the unmet need.”

The opportunities extend far beyond health care. Research firm IDC estimates the commercial sector will generate 5.8 zettabytes of productivity data—sales forecasts, customer data, and the like—this year. A zettabyte of information corresponds roughly to the number of grains of sand on all the world’s beaches. A lot, in other words, says John Rydning, head of IDC’s Global DataSphere program, which measures the amount of data created each year. 

5.8 zettabytes


That’s the amount of productivity data that research firm IDC predicts will be created this year, presenting an opportunity for A.I. to crunch that information. A zettabyte is roughly equivalent to the number of grains of sand on all the world’s beaches.

This means that businesses of all types, if they can corral the data into a form neural networks can learn from, have a lucrative opportunity. Even slight improvements in predictive capabilities can lead to enormous financial gains, says Athina Kanioura, chief strategy and transformation officer for food giant PepsiCo. “The additional level of accuracy translates to millions of dollars,” she says.

The challenge, then, is getting researchers to work with the kind of data that can be most helpful to business. “The deep networks that are so cool can really do amazing things for our cars and for understanding sentiment from tweets online,” says Peter Bailis, a Stanford professor and also CEO of a Silicon Valley startup called Sisu Data that builds analytical tools for businesses. “But they don’t help us with understanding things like risk or customer satisfaction if our data is stored in tables.” In terms any businessperson can relate to, the question remains: Can A.I. conquer its Excel problem?

Newsletter-Red-Line-15

Progress in promoting business applications for neural networks rests on getting the programs to understand words as well as they have been able to analyze images. For that, researchers have turned to a technique called word2vec. (The “vec” stands for vector, the type of analytical unit best understood by a neural network.) Word2vec, invented in 2013 by a team of Google researchers and published as an open-source software project, helps computers map the relationships among certain words. It has led to more powerful language systems that recognize, for example, that the word “car” is more closely related to automakers like BMW or Nissan than a food company like Kraft Heinz.

The computational magic of word2vec is its ability to discover those correlations by converting words into a string of numbers that neural networks can understand. Over time, as a neural network is trained on additional text, it groups words according to numerical scores measuring how frequently the words appear near each other. Compared with older so-called natural language processing technologies, these newer systems improve on the pattern recognition attributes typically associated with human thought. 

From this computer-assisted word-association game comes an ability to make sense of what is stored in the rows and columns, for instance, of a spreadsheet. This process creates a type of Morse code for a neural network: If the program comes across a sales spreadsheet with a column indicating “days,” it can learn with enough data that certain holidays could impact sales during a particular season without being explicitly told to do so. “It’s kind of the core idea,” says Rachel Thomas, director of the University of San Francisco’s Center for Applied Data Ethics and cofounder of an educational nonprofit called Fast.ai. “Neural networks are providing this infinitely flexible architecture for learning by modeling a particular shape of patterns.”

What makes artificial intelligence look dumb?
Former Uber engineer Rachel Thomas cofounded the educational nonprofit Fast.ai and an ethics focused think tank in San Francisco. She explains A.I. to businesspeople and scientists alike.
Gabriela Hasbun

The investment world alone is rife with opportunities for analyzing words. At Goldman Sachs, a team of researchers trained a neural network to look for words associated with intra-family home transfers. Such noncommercial transactions likely won’t describe the true value of a house, and teaching a software program to factor them out can improve the bank’s analysis. “So we trained a neural network so it learns to pay less attention to a transaction that has that label,” says Charles Elkan, a longtime professor of computer science at the University of California at San Diego who until recently led machine learning projects for Goldman. 

Sophisticated word association is also invaluable for logistics operators. The San Francisco grocery-delivery startup Instacart uses a variant of word2vec to teach its algorithms to anticipate customer preferences, particularly when requested items are unavailable. The program converts the words for supermarket inventory items into numerical data so neural networks can process them. The network then groups items together so it can understand, for example, that trail mix has more in common with dried fruit or nuts than it does with coffee. The result is a time and money saver, says Sharath Rao, a machine learning director for Instacart. “Otherwise you would have to think of all the possible pairs and keep a [manual] table,” he says. 

Newsletter-Red-Line-15

For all the momentum behind using deep learning on structured data, hurdles remain. For one, the idea is so new that there’s no tried-and-true way to evaluate how good these techniques are compared with more conventional statistical methods. “It’s a bit of an open question right now,” says Even Oldridge, a data scientist for Nvidia, which makes chips that power A.I. software. 

Indeed, given the expense of training neural networks, older data analytics methods may be sufficient for companies that don’t have the right A.I. expertise in-house. “I’m a firm believer that for every company, there isn’t a magic solution that can solve every problem,” says A.I. expert Kanioura, the PepsiCo executive. This is in fact behind the pitch that cloud-services giants Amazon, Microsoft, and Google make: Buy A.I. services from us rather than making large expenditures on talent for potentially incremental returns. 

It’s the core idea. Neural networks are providing this infinitely flexible architecture for learning by modeling a particular shape of patterns.

Rachel Thomas, cofounder, Fast.ai

And as with any project where humans aim to teach computers how to “think,” the biases of the living organisms threaten the project. Deep learning systems are only as good as the data they are trained on, and too much or too little of a certain data point can skew the software’s predictions. Genentech’s data set, for instance, has clinical data on cancer patients dating back 15 years. However, the genomic testing data it uses in its spreadsheet is eight years old, meaning that patient data from before then isn’t as comparable as researchers might like. “If we don’t understand these data sets, we could build models that are totally unreliable,” says Genentech’s Copping.

Still, the potential value of supercharging the analysis of all those spreadsheet fields is nothing less than being able to “predict how long a patient can survive” with a certain treatment, says Copping. Not bad for a bunch of rows and columns.

Newsletter-Red-Line-15

A handful of corporations are teaching neural networks to work with the kind of structured data that already exists within their walls. A few examples:

Genentech

The biotech pioneer has built a spreadsheet with complex health data from routine records to genomic profiles—from tens of thousands of patients. The stakes are high: If artificial intelligence can properly analyze the data, the result could be medical treatments targeting the disease of iindividual patients.

Goldman Sachs

A.I. presents untold opportunities for investors. The bank hired a machine learning professor to build a tool to teach networks to ignore phrases that could complicate a financial analysis. Example: “Intra-family transfers” likely don’t reflect the accurate value of a home. Teaching a network to find them can improve the model.

Instacart

The grocery-delivery startup has an understandable data set in the inventory of supermarket items its workers pick for customers. The company is teaching its algorithms to do sophisticated word association like matching trail mix with nuts and dried fruit—in order to offer customers alternatives when their choices are out of stock.

A version of this article appears in the October 2020 issue of Fortune with the headline “What makes artificial intelligence look dumb.”