Skip to main content

IBM R&D Labs in Israel News

ICML and COLT 2010

Ask me about analytics
Humans learn many things for years and become better learners over time. Why not machines?

  An ICML poster session
What would it be like to build a system that learned forever, cumulatively, over many years? This was the question asked by keynote speaker Professor Tom Mitchell at the 27th International Machine Learning Conference (ICML) hosted by IBM Research in Haifa, Israel. Tom’s group is attempting to build a never-ending learner for language, and his talk was just one of the fascinating lectures and workshops featured at ICML 2010 and the co-located Conference on Learning Theory (COLT 2010).

Adding intelligence to systems

Although machine learning has been around for a while, the field is now seeing more and more practical applications being fielded and more people participating in the field. As explained by Mitchell, by following certain rules, a system can detect patterns and begin to ‘understand’ the meaning of certain words. For example, Mitchell’s team is trying to get machines to crawl through masses of text and categorize or classify nouns. If a noun starts with a capital letter and often occurs together with phrases such as ‘mayor of…’ or ‘lives in …’, then it can be assumed to be the name of a city. If a noun appears together with the term "Ltd." or "Inc.", then we can ‘learn’ that it’s the name of a company.

"Hosting the ICML and
COLT conferences is
similar to hosting the
Olympics, with the same
desire to make it an
unforgettable experience
for everyone"

But what happens when the machine encounters a phrase like “Alice lives in denial”? The machine will continue learning and may ‘understand’ that certain human emotions like denial or humiliation can be classified as cities. This is just one of the challenges being addressed by experts like Mitchell who want to overcome the learning plateaus faced by learning systems through better formulation of problems or the coupling of different techniques for constraining learning.

Other conference highlights included the ICML talk by Nobel Laureate Prof. Robert Aumann on perfect information game theory, the COLT lecture by Prof. Noga Alon on the impossibility of fair voting schemes, and a grand conference banquet in the ancient city of Acre.

The Olympics of machine learning

  IBM researcher Dan Pelleg
The two conferences hosted over 600 participants from 26 different countries. ICML had close to 600 papers submitted with 152 accepted, 164 lectures in 5 parallel tracks, 9 workshops, 7 tutorials, 3 locations, and 1 Nobel laureate. COLT had nearly 130 papers submitted with 41 papers accepted, and six open problem sessions and two invited talks. This is also the first time ICML has been held in the Middle East or Asia. Whether or not the machines can learn from it, these numbers are testament to the importance of the field and the talent of the local team at IBM Research – Haifa who hosted the conference.

“Hosting the ICML and COLT conferences is similar to hosting the Olympics, with the same desire to make it an unforgettable experience for everyone,” explained Dr. Shai Fine, one of the local chairs for the conference and manager of the Analytics Department at IBM Research – Haifa. “Being able to host the conferences is both an honor and a privilege—as well as a statement of support for the quality of research at our lab,” he said. As noted by many conference-goers, the local organizers did a fabulous job with everything from hotels, meals, wifi, and transportation through tours, meetings, T-shirts, web sites, and graphic design.


A key to building a smarter planet

   At the ICML conference banquet
Machine learning is key to IBM’s drive to build a smarter planet, helping add intelligence into the systems and processes that make the world work. This is being done by gathering information from trillions of sensors and digital devices, connected through the Internet. Machine learning and advanced analytics provide the computational power to make sense of all this data and turn it into knowledge. This same wisdom can help us reduce costs, cut waste, improve productivity, make things run smoother, and even transform the way an entire industry does business.

“With so many opportunities and areas for insight, it’s no longer possible to have specialized experts handcraft solutions for each individual problem,” explained Dr. Dan Pelleg, senior researcher for analytics at IBM Research - Haifa, and local chair for the conference. Machine learning is like a secret ingredient that can be added to help solve problems, even without specialized, deep-rooted expertise in a particular domain. As Pelleg sees it, the days of the lone inventor sitting in an isolated room are over. “It’s the teamwork between the analytics experts and the subject matter experts that becomes most important in these cases," he emphasized.

Collaboration was definitely a high priority for everyone at ICML and COLT. Walking through the conference venue and hotel lobbies, small groups of colleagues gesturing, huddling, and debating over various solutions was a common and welcome site. “Thank you so much for the great work on this conference. It was an unforgettable moment from my perspective,” noted Martial Hue, a researcher from Mines Paristech.

Analytics in Haifa

Over the years, the machine learning group in Haifa has added ‘smarts’ to many vital solution areas, including:

  • Improving coverage directed test generation
  • Predicting churn for telco operators
  • Helping physicians determine the optimal treatment for HIV patients
  • Improving yields in IBM’s fab plants
  • Conserving water through insight from water meter data
  • Detecting fraud in financial transactions
  • Extracting information on system failures from masses of data in log files
  • Optimizing system backup and recovery
  • Improving manufacturing
  • Understanding the correlation between genetic variations and different illnesses

ICML and COLT web sites

Click on the logos of the ICML and COLT conferences below to learn more about these two influential conferences.

ICML logo COLT logo

The Haifa Analytics Department

The IBM Research - Haifa Analytics Department focuses on research and development of machine learning techniques and constraint satisfaction problem (CSP) algorithms. The group applies its expertise to problems in various areas, such as information retrieval, autonomic computing, anomaly detection, biological and medical mining, and workforce management, and provides solutions, tools, core technologies, and services.