Neural network learns to count - by the way
Many vertebrates are capable of rudimentary counting. For example, you can choose a point pattern that contains more or as many points as another-at least after appropriate training. As it turns out, this may not even require a special sense of numbers: It could already be enough to learn to recognize objects in general, with the understanding of the size of sets being a by-product.
This is suggested by a work by scientists led by Andreas Nieder from the University of Tübingen. They trained a neural network to recognize objects in photos. They then discovered that their network was also able to classify objects by number - without ever having been trained to do so.
Nieder and colleagues used a deep learning network in their model. Such systems are surprisingly good at recognizing objects in images if you show them enough of them - over a million photos with correct answers, for example. In the course of the learning process, the artificial neurons take on the role of feature detectors. They specialize in recognizing certain image properties such as shapes, edges and brightness distributions. At the same time, the network learns which image properties typically come together when a displayed image is classified as a tennis racket, schnauzer or crane, for example.
This is how the researchers trained their network until it was sufficiently good at recognition. Then, to test their network's mathematical abilities, they gave it patterns of points to analyze - black boxes of 1 to 30 points of different shapes and sizes - and looked at which feature detectors deep in the network "triggered". In fact, the statistics told them that among the many thousands of artificial neurons, some apparently specialized in displaying the sheer number of objects.
The "quantity detectors" behaved in a way that is familiar from similar cells in the primate brain. For example, a neuron that is specialized for the number six also becomes active when five or seven objects are presented, but with less intensity. In addition, networks - both natural and artificial - find it easier to distinguish between small numbers than large ones. From the point of view of the researchers, these similarities in behavior suggest that both systems operate according to the same principles.
The spontaneously developed understanding of numbers is not real counting, the researchers explain. Your network has learned to deal with sets where all elements are shown simultaneously and side by side. When counting, however, it is important to consider a sequence of numbers. Which skills are necessary for this and whether these can also arise spontaneously in neural networks must be clarified in future studies.