In the past decade, computer vision algorithms have come a long way. They have been shown as good as people at tasks like categorizing cat breeds or dog breeds and they have significant ability to identify specific faces among the millions. Researchers at Brown University shows that computers fail at a class task that young children faced no problem and determine it whether two objects in the image are the same or different. In the last week, paper presented at the annual meeting of Cognitive science society, the brown group sheds light on why computers are bad at these kinds of tasks and suggest avenues towards the more vision system smart computer.
Thomas Serre, associate professor of cognitive, linguistic and psychological sciences at Brown and the paper’s senior author said that there are various excitements about the computer vision and it has to be achieved and he shares a lot of that. He said that they are working with the current computer vision limitations and they have done here. They can move toward new and more improved systems rather than taking out the system, which they have already.
Serre and his colleagues used state-of-the-art computer vision algorithms to analyze black and white images containing two or more randomly selected shapes. Some cases the objects are identical and they were completely different. The computer asked to identify the same or different relationship.
The problem source in individuating objects is the machine’s architecture learning system that powers the algorithms. The algorithms use neural networks. Serre says that while its unclear about the feedbacks, that they have something to do with their ability to pay attention to some parts of their visual field and make mental objects representation in their minds, that means people attend to one object. This builds a feature representation that is bound to the object in their working memory. Then, they shift their attention to another project. While both the objects are representing in the working memory, your visual will make the comparison to some or to other.
Serre and his colleagues give the reason that the computers can’t do anything, because of its feed-forward neural networks and this doesn’t allow the type of recurrent processing needed for this mental representation and individuation of objects. Serre says that making computer vision smarter, neural networks would require that closely approximate the nature of human processing that can be visualized.