Introduction
Welcome to the Embracing Uncertainty web site, an accompaniment to our exhibit at the 2010 Royal Society Summer Science Exhibition which was held in London from June 25 to July 4 2010. Here you will find information about the various demonstrations which formed part of our exhibit, and you can even try out one of the demonstrations for yourself within your web browser.
Embracing Uncertainty
Scientists have worked for decades to try to create intelligence in computers. Traditional approaches relied on hand-crafted solutions and had limited applicability. New techniques being developed at Microsoft Research Cambridge and other institutions around the world, are based on computers which can learn for themselves by analysing large sets of data. This capability is known as machine learning.
The key to learning from examples is to recognise that real-world data is full of complexity, ambiguity, and uncertainty. Computers can be programmed to handle these challenges by using a branch of mathematics called probability theory. The short video on the home page of this web site includes a demonstration of a “Galton machine” which illustrates how probability theory allows random events to be described in a precise way. In the case of the Galton machine, the random events correspond to beads bouncing randomly off the pins inside the machine. Visitors to the exhibit had the opportunity to try the Galton machine for themselves.
By embracing the mathematics of uncertainty, we are able to create machines that can learn from data. The Clinical Drug Trial demonstration is a simple example which illustrates the use of probabilities to model the results of a trial for a new drug. We see how the probabilities change as we show more data to the computer. Such changes in the probabilities correspond to a reduction in uncertainty as the computer learns from the data.
The Movie Recommender demonstration provides a practical example of how probabilities and machine learning can be applied to the problem of recommending movies to people based on ratings they have given to other movies which they have already seen. This data is combined with the recommendations made by lots of other people, and allows the computer to work out, for any new movie, the probability that the user will like it.
LiveObject illustrates the use of machine learning for the visual recognition of everyday objects such as pens and mobile phones. The system can be trained on a new category using a few example objects and the accuracy of the system improves as it sees more examples of that object.
OrganNavigator is a practical application of recognition technology, which uses probabilistic methods to determine the position and extent of anatomical structures in conventional 3-dimensional medical scans.
Finally, Background Removal shows how machine learning can be used to simplify tasks in image editing. Using just a few mouse clicks, the system removes the background from an object (such as an animal or a person) within an image, allowing the object to be pasted into a new image.
Visitors to the exhibit were able to explore the science of uncertainty for themselves through hands-on interactions with all of these demonstrations. Those wishing to delve more deeply into the science and mathematics behind our exhibit can look at the section below on The New Machine Intelligence.
The New Machine Intelligence
We are at the start of a revolution in computing driven by data, and enabled by the advent of cloud computing. In recent years, a new framework for machine intelligence has been developed which can drive and exploit this revolution, and is based on three key ideas.
The first is the use of probabilities to model and quantify uncertainty. This is known as the Bayesian view of probability, and has its origins in a seminal paper published by the Reverend Thomas Bayes in 1763. This paper is one of 60 publications showcased by the Royal Society in their interactive Trailblazing timeline.
The second key idea also has a long history and involves the use of graphs to describe the relationships between the various quantities that are involved in a particular application. These probabilistic graphical models offer a powerful approach to the design of a machine intelligence solution.
Finally, the new ingredient in this framework is a set of very efficient algorithms for computing the required outputs of the system, once the training data has been collected. These algorithms have been developed over the last five years and they have an elegant interpretation as local “messages” passed around the graphical model. Their efficiency allows Bayesian methods to be scaled up to massive applications involving millions, or even billions of data points.
Examples of large-scale applications of this new framework include the TrueSkill™ Bayesian skill rating system which is widely deployed on Xbox Live®, AdPredictor for predicting click-through rates in web search and the Matchbox collaborative filtering and recommendation system.
A sophisticated software package, called Infer.NET, which implements this new framework for machine learning, has been made available as a free download for non-commercial use.
Visitors who would like to learn more about the exciting developments in machine learning can watch the 2010 Turing Lecture Embracing Uncertainty: the New Machine Intelligence, while those wishing to study the detailed mathematics on which it is based are recommended to look at the text book Pattern Recognition and Machine Learning.
