In his book Outliers, psychology writer Malcolm Gladwell says that achieving greatness in a field of human endeavor often rests on two essential factors: inherent talent, and many hours of deliberate practice at a young age. Cornell Associate Professor of Computer Science Kilian Weinberger began working on the second factor very early, writing his first computer program at the age of seven. “I grew up near Munich, Germany and I spent at least 10,000 hours on a computer before I even started college,” says Weinberger.
At nineteen, Weinberger applied to the University of Oxford in England and received his undergraduate degree in mathematics and computer science in 2002. He then went on to the University of Pennsylvania, where he earned a Master’s in Computer Science and a Doctorate in Machine Learning. Weinberger’s advisor at UPenn was Lawrence Saul. “I was one of Lawrence’s first grad students, and his lab was full of energy. His enthusiasm for machine learning was contagious.”
Weinberger then worked for two years as a Research Scientist at Yahoo! Research. “These days everyone is excited about ‘Big Data’. But back in 2007, Yahoo! was one of the only companies that recognized the value of large data sets,” says Weinberger. “In Yahoo!’s research lab, we had access to large-scale social networks, webpages from the entire internet, logs of web searches, and millions of images. I realized the widespread impact that machine learning and data sciences would have in the near future.”
The sorts of problems that Yahoo! engineers would bring to Weinberger all involved the need for powerful learning algorithms. “Applications engineers are really great at what they do, but for some tasks the code simply cannot be written,” says Weinberger. “That is where learning algorithms come in.” The example Weinberger gives to make his point clear is this: How would you write a program to ensure that the computer could recognize a picture of your grandmother?
After just a moment’s thought, the difficulties of the problem become clear. Profile or head-on? Bright light or dim? With her hair straight or in curlers? At age 55 or age 82? The amount of code a programmer would need to write would be impractically large for just this one task. If instead you give a computer a learning algorithm and then you present it with a very large number of pictures and answers to the question “is this your grandmother?” the computer develops a sharper and sharper idea of what your grandmother looks like. With enough data, it can become very good at identifying pictures of your grandmother. The computer can be said to have “learned” the task from examples.
This idea of machine learning is at the root of Weinberger’s current lines of research at Cornell’s Department of Computer Science.
One area of interest for Weinberger is Resource-Efficient Machine Learning. Essential features of digital life such as search engines and spam filters are algorithms. Weinberger is exploring how to maximize the efficiency of these algorithms given resource constraints like processing speed and power, battery limits, and timeliness. For example, Weinberger and his students have found that one way to cut the resources needed to run an algorithm is to compress it to a small fraction of its original size, much like an image can be compressed into a jpeg without losing its essential characteristics. He sees the realm of resource-efficient machine learning as one that could lead to a surge in improvements across many high-impact application domains.
Another line of research Weinberger is pursuing is making algorithms interpretable. The example he gives is a patient in an Intensive Care Unit at a hospital. The patient is connected to many sensors and a variety of data is being collected in real time about blood oxygenation, heart rate, body temperature, brain activity, and many other measureable parameters. This constellation of numbers can be run through an algorithm to recognize signs of an impending medical crisis. When implemented in a real medical setting, however, these early warning systems were often ignored (even when they were correct) because the attending doctor was not given a clear explanation for the alert. Weinberger says that making the algorithm correct is not enough. “When the system produces a warning,” says Weinberger, “it also needs to provide an actionable explanation to the medical practitioner.”
Other areas of interest to Weinberger are metric learning, machine-learned web-search ranking, transfer- and multi-task learning, and biomedical applications. Before coming to Cornell, Weinberger taught for five years at Washington University in St. Louis. While there he collaborated with researchers from their renowned medical school on projects involving nanoparticles, gene expression, and decoding brain signals.
Weinberger enjoys classroom teaching very much and is looking forward to the Machine Learning class he will be teaching this fall. “I’ve always loved teaching. When I was at Yahoo! Research I enjoyed working with interns, and working with students was one the best parts of my job at Washington University,” says Weinberger. “I can’t wait to get to know the students at Cornell.”
When Daniele Visioni was six months away from finishing his physics Ph.D. studies at Italy’s University of L’Aquila he took a long bus ride to Zurich, Switzerland to give a talk at an academic... Read more about Assessing the feasibility (and advisability) of using Sunlight Reflection Methods
Back in 2008 when EAS Assistant Professor Riley Culberg was an undergraduate at the United States Military Academy at West Point, it is unlikely he would have predicted he would end up on the faculty... Read more about Using radar to figure out where glacial meltwater goes
When Grace Barcheck was a sophomore at Washington University in St. Louis, she took a geoscience class that changed her life. You can trace a direct line from that class to her current position as one... Read more about Studying plate tectonics with new tools--and a new perspective
