Machine learning and natural language processing - The Computing Universe: A Journey through a Revolution - page 287

Research, Heckerman has applied Bayesian techniques to problems such as
spam detection in email and troubleshooting failures in computing systems.
He now leads a research team analyzing genetic data to better understand the
causes of diseases such as HIV/AIDS and diabetes. His team recently performed
an examination of the genomes (complete sets of DNA) of many people to
find the genetic variants associated with particular diseases, using data from
a Wellcome Trust study of the British population. For each of seven major
diseases, the data includes genetic information about two thousand individu-
als with that disease. The data also include similar information for thirteen
thousand individuals without any of the diseases. Using a new, computation-
ally efficient algorithm that Heckerman's team has developed to remove false
correlations, the researchers analyzed 63,524,915,020 pairs of genetic markers
looking for interactions among these markers for bipolar disease, coronary
artery disease, hypertension, inflammatory bowel disease (Crohn's disease),
rheumatoid arthritis, and type I and type II diabetes. They processed the data
from this study using twenty-seven thousand computers in the Microsoft cloud
computing platform, an Internet-based service in which large numbers of pro-
cessors located in a data center can be used on a pay-as-go basis. The comput-
ers ran for seventy-two hours and completed one million tasks, the equivalent
of approximately 1.9 million computer hours. If the same computation had
run on a typical desktop computer, the analysis would have taken twenty-five
years to complete. The result was the discovery of new associations between
the genome and these diseases, discoveries that could lead to breakthroughs
in prevention and treatment.
Computer vision and machine learning:
A state-of-the-art application
Humans find vision easy and can look at a scene and rapidly understand the
objects in the scene and the context in which these objects coexist. Computer
vision still has a long way to go to match human vision even though this has
been a key research area in computer science since the mid-1960s. Although
progress has been slow, there are now many commercial applications of com-
puter vision algorithms, ranging from industrial inspection systems to license-
plate number recognition. In the early 1990s, computer scientists developed
vision-based systems to capture three-dimensional human motions. One such
system could recover the three-dimensional body positions of a person moving
in a special studio wearing clothing with special reflective markers, by collect-
ing images from multiple cameras. Research also continued on algorithms that
could recover three-dimensional information from video footage. However, the
problems of image understanding and general object recognition remain huge
challenges for computer science. Some progress has been made ( Fig. 14.3 ), but
for major advances to occur, software models of each object need to be gen-
erated from the data rather than handcrafted by the programmer. Machine
learning is now recognized to be a key technology for effective object recogni-
tion. In 2001, Paul Viola and Michael Jones used machine-learning technologies
to build the first object-detection framework that could provide useful detec-
tion accuracy for a variety of features. Although their system could be trained
to recognize different classes of objects, they were motivated to design their