In the field of artificial intelligence, Ken Forbus and team of researchers in Northwestern University, U.S., have found a way to make a computer learn through analogy and have the ability to reason and make decisions by using analogies, in much the same way a human does.
The new additions to the Structure Mapping Engine model makes it possible to scale up to more complex analogies and engage with complex functions such as visual perception and tackling moral dilemmas. The research is published in the journal Cognitive Science.
The latest work on SME adds five extensions that were needed to model analogy in large-scale tasks. “These extensions have also enabled us to use analogy in a sketch-based intelligent tutoring system, CogSketch, which is freely available on the web,” said Dr Forbus in an email to this correspondent.
The model relies on the structure mapping theory of analogy and similarity developed by psychologist Dedre Gentner of Northwestern University. Earlier models of SME and those using analogies in computer learning were unable to tackle the size of complexity demanded by human activity.
Analogies can be simple or complex, and humans use complex analogies to help in decision-making. The improved idea of the structure-mapping engine can handle the size and complexity of relational representations that are needed for visual reasoning, tackling textbook problems and moral dilemmas.
“Conceptually, SME is very simple, but there are ample subtleties in the implementation. That is why we made source code publicly available at the same time, so that other researchers can start with it,” says Dr. Forbus.
The SME way of using analogies, needs fewer examples to learn from; however, the use of analogies needs to be guided by a feeling for the way humans produce knowledge and how they relate different bodies of knowledge.
“...The systems which produce representations need to produce structured, relational representations. Such relationships seem to be an essential part of human knowledge: We think about relationships between objects, people and ideas, we plan and explain and construct theories. But many of today’s artificial intelligence systems only use vectors of features, which are not expressive enough to capture relationships at scale,” he adds.
Developing artificial intelligence can be done in two ways — by emulating humans and by trying a totally different tack. Other systems of artificial intelligence use methods such as deep learning, which in turn make use of the power of computers to handle huge sets of data and learn by sifting through huge data banks.
It is a moot point whether simulating human thinking is the best way to create artificial intelligence. There are instances of the success of both, namely, imitating nature and modelling something totally at variance, in the history of technology. For instance, humans had tried unsuccessfully for years to emulate nature for the purpose of flying, by imitating birds.
However, aircraft which work on a completely different principle were the first to take off.
In the field of artificial intelligence, deep learning and SME mark these two different paths to artificial intelligence. Deep learning is a non-human way of modelling artificial intelligence, whereas SME follows the more complex, humanlike, route of analogies.
