The central advancement with artificial intelligence is improvement by learning. The more AI understands when something is either right or wrong, the better predictions and other data inferences become. But what about creativity, or at least dealing with the unexpected? How is AI to learn to expect and address the unexpected?
The answer may be with a form of distraction. This comes from researchers based in the Netherlands and Spain, who have determined how a deep learning system designed for image recognition can more readily move through learning process when it has an additional need to be concerned about secondary characteristics. By this, the reference is with data that is not directly related to the matter at hand.
The technology being evaluated by the University of Groningen scientists was to do with image recognition.
The basis of the technology is a type of Convolutional Neural Network. This is a form of bio-inspired deep learning in artificial intelligence. The concept involves the interaction of thousands of ‘neurons’, intended to mimic the way the human brain learns to recognize images.
While technologists can demonstrate that the networks are successful, it remains unclear precisely how they work. However, what the research team have established is that for image recognition sometimes trivial details are just as important as the main object.
The scientists have been considering this in relation to teaching an AI system to recognize different types of food. The way around this an to prove image detectability was found to be adding in routines that deliberately and periodically distract AI systems from their primary targets.
In other words, making the system to look elsewhere for identifiers beyond the core activity. Through such application it was found, with the food recognition studies, that by using extra information, the AI learnt better and the system became more fine-grained in its classification of objects.
Image recognition is important for making AI, and robotics, better. The classical problem in computer image processing is that of determining whether or not the image data contains some specific object, feature, or activity. The new approach offers a means to make this task easier.
The research appears in the journal Neural Computing and Applications. The research paper is titled “Playing to distraction: towards a robust training of CNN classifiers through visual explanation techniques.”