Today’s computers are doing a lot more of our thinking for us than we realize, says Carnegie Mellon University professor Tom Mitchell.
When we swipe our credit card at a store, a computer program decides whether it’s a valid purchase. When we get an email, a computer program decides if it’s spam. And when we go on Facebook, a computer program decides which ads and news posts we see. Increasingly, these programs are no longer based on following simple rules, but instead continuously learn from real-life experience and change their operations as they go along.
The special type of software that makes these decisions is known as a machine learning program, and it’s an area that Mr. Mitchell has pioneered at CMU. In a special section on artificial intelligence in this week’s Science journal, Mr. Mitchell and University of California-Berkeley professor Michael Jordan describe the state of the art in machine learning and speculate about where the field is headed.
Many of the commercial programs in use today are examples of “supervised” machine learning, he said in an interview this week.
That means that the programs train themselves on a database with known results, and then use that to create a model of what is happening with new data.
One example in the Pittsburgh area is an apartment rental website known as EDigs, started by CMU professor Jen Mankoff and her colleagues.
EDigs helps prospective renters predict what their utility costs will be, based on information about the apartment, as found in an ad, and details about themselves, she said.
The EDigs machine learning program trained itself on a huge government database known as the U.S. Residential Energy Consumption Survey, which includes information on energy use in about 12,000 households. Using a method known as regression analysis, it figured out which residential and lifestyle factors were most important in determining electricity and natural gas use.
Once a person goes on the EDigs website, she can fine-tune the utility prediction by answering more specific questions, such as what indoor temperature she prefers or how many TVs might be operating simultaneously.
Another supervised machine learning program is the foundation for a San Francisco company known as Sift Science, which helps online businesses detect fraudulent transactions.
Doug Beeferman, who graduated from Carnegie Mellon in 1998, is a software engineer there. The company’s software looks for patterns that predict such transgressions as credit card fraud, fake product reviews or spamming of other customers. When the company was getting underway in 2011, he said, it was given access to a database from the vacation rental site Airbnb.
That allowed the software to train itself on real-life examples of legitimate and fraudulent behaviors. It then applied those lessons to information from new customers, and the more examples it has accumulated, the better it has become at detecting patterns of abuse, he said.
By seeing how well its predictions match up with actual results, the program can continually fine-tune itself, Mr. Beeferman said, and in that sense, it operates much like the human brain does.
The self-educating nature of machine learning shows up in many other applications, said Mr. Mitchell, a Ph.D. who is chairman of CMU’s machine learning department.
For instance, experimental vehicles often use machine learning software to improve their ability to identify the objects detected by their cameras. “If a vehicle’s camera looks and sees a spot on the road 100 yards away and can’t identify it, as it is driving down the road, at some point it will be able to see that it’s a bike. And now it can go back and retrieve that 100-yard image and can label the image as an example of what a bicycle looks like 100 yards away.”
So, what makes these machine learning programs any better than a group of smart humans looking for the same patterns?
For one thing, the computer programs don’t carry any biases into their work, said Siftscience’s Mr. Beeferman.
In the field of credit card fraud, he said, some companies have banned all credit card transactions from Nigeria.
“But really the data says that in fact the great majority of Nigerian credit card shoppers are good users,” and the computer program knows that it’s more important to look at such factors as whether the credit card billing address is from the same country as the user, or what that card’s recent purchase history has been.
While supervised machine learning programs are popular in the commercial world, many research programs use machine learning software that hunts for patterns in data it has never seen before.
That’s the kind of program Mr. Mitchell and CMU psychology professor Marcel Just have used in their groundbreaking work to predict what words or phrases people are thinking of merely by looking at their brain activation patterns in a functional magnetic resonance imaging machine.
The software is able to see which regions of the brain are most active when people look at certain words, and then use that information not only to guess what word someone is thinking of, but to predict what the brain activity pattern would look like for a brand new word. The machine learning algorithms have shed new light on how human brains function by showing that the activity pattern for any word combines different functional areas of the brain. So the pattern for the word “watermelon” includes activity in the brain’s taste centers as well as in the tactile center, which might relate to the watermelon’s smooth oblong bulk, he said.
In a similar way, machine learning programs can find meaningful clusters of information in huge databases such as the genomes of thousands of people or the text from millions of documents.
In their Science article, Mr. Mitchell and Mr. Jordan say one new frontier for machine learning might be to combine the analytical power of the computer with the diverse knowledge base of human experts.
A machine learning program can look at a medical database and tell which treatments work best for which kinds of patients, Mr. Mitchell said, but it is limited to whatever information is already in the database. If the patients’ doctors could then look at the results, he said, “they may know that some of them had bloodshot eyes, or that half of them had not reached puberty,” and that information can help refine the diagnosis.
Another future focus he and Mr. Jordan identify are “never-ending learning programs” that might operate for years at a time. “Right now 99 percent of machine learning algorithms operate on a fixed data set and then we turn them off,” Mr. Mitchell said, but they could be built to function more like human beings, who constantly build on previous knowledge to move themselves to the next level.
“My real-dream scenario,” he said, “is to have a never-ending learning system that is my assistant for my research in neuroscience, which would help me try to learn what each part of the brain is doing. I want a program that is not reading just the Web, but also the research literature in neuroscience and that would develop rules for how to interpret that, and would look at experimental brain imaging data as well.”
Mark Roth: mroth@post-gazette.com, 412-263-1130 or on Twitter @markomar
First Published: July 16, 2015, 6:30 p.m.
