On a mission to improve man-machine collaboration, researchers at MIT have developed a new computer learning model which helps computers to learn rather like human beings do, i.e. by drawing on examples of similar scenarios.

Computers begin to learn like people

The Massachusetts Institute of Technology (MIT) researchers behind the new machine learning model have based their work on two parallel observations: computers are very good at identifying patterns that come from bulk data sets, while humans are gifted when it comes to understanding patterns sourced from just a few examples. The challenge for researchers Julie Shah, Been Kim and Cynthia Rudin was therefore to find a means of bridging these two ways of processing information in order to improve the way human beings and computers work together on decision-making. Julie Shah, assistant professor of aeronautics and astronautics at MIT, who co-authored the recently-published paper on the work – ‘The Bayesian Case Model: A Generative Approach for Case-Based Reasoning and Prototype Classification’ – uses a practical example to explain what she and her colleagues are doing:  “It’s the type of decision-making people do when they make tactical decisions, like in fire crews or field operations. When they’re presented with a new scenario, they don’t do search the way machines do. They try to match their current scenario with examples from their previous experience, and then they think, ‘OK, that worked in a previous scenario,’ and they adapt it to the new scenario.” The MIT team set out to apply this very human thinking process to computer information search.

Setting the ‘prototype’ shapes the decision-making process

The new model – called the Bayesian Case Model (BCM) – put forward by the MIT trio is in fact fairly simple. They worked on a type of machine learning known as ‘unsupervised’, where the computer looks for commonalities in unstructured data. The result is a set of data clusters whose members are in some way related, but it may not initially be obvious how, as the computer does not work from set labels, as is the case with ‘supervised’ machine learning. The researchers made two major modifications to the type of algorithm commonly used in unsupervised learning. The first was that the clustering was based not only on the shared features of data items, but also on their similarity to a representative example, which the researchers have dubbed a ‘prototype.’ The other is that rather than simply ranking shared features according to importance, the new algorithm tries to reduce the list of features down to a representative set, which the researchers are calling a ‘subspace’. When tested out on a range of classic machine-learning tasks, the new algorithm performed as well as its predecessor on most tasks, actually performing better on some of them.

Applications in robotics and other fields

So the major question going forward will be to see whether the BCM will improve computers’ or indeed humans’ ability to accomplish certain tasks. This is what Julie Shah’s team is now working on. So far they have been testing the model on data sets such as food recipes. The model categorises them based on their most prominent ingredients, plus their similarity to a representative example (‘prototype’) for any given cluster of recipes, which is also chosen by the computer. Using the BCM approach achieved results that were 15% better than those attained with previous models. The basic goal of this research was to improve the outcome when human beings draw on the results worked out by computers for their decision making, with each learning from the other. However, the new model looks set to  be a boon for robotics as well. Ashutosh Saxena, currently visiting professor at Stanford University working with his group on the Robo Brain project – whose purpose is to improve how robots learn using the Internet – has expressed his great interest in the new BCM model. Following recent work to develop machines that can understand our emotions and to build computers which teach each other skills,  this most recent research seems likely to take machine capabilities even further, though with the same basic reference point – the human user – in mind.

By Guillaume Scifo