2017-10-15 10:59:02
The development of intelligent artificial vision systems has made it possible to compare the performance of humans and robots in the tasks of identifying objects in a complex scene. Where is the tomato jar in the fridge? Where is my car in that photo of the parking lot they show me? Studies have revealed that animals, from insects to humans, learn about the probabilistic and statistical relationships in their environment to guide their visual system toward the correct identification of a given object. For example, if volunteers are asked to find a toothbrush in a bathroom, they usually start by looking around the sink, not in the back of the toilet. Similarly, if I want to find the tomato jar in the fridge, I will quickly discard square or other shaped objects that deviate from the expected cylindrical shape for a jar, which will help me identify and find it more easily. Thus, the relationships, among others, of the expected shape and position of a given object are used by our brains to identify it.
Studies on visual ability carried out with volunteers have revealed that if the object is placed in an unusual position in the scene, it is always more difficult to find it. What happens if the tomato jar is made the size of a small salt cellar, or if the toothbrush grows as big as a toilet brush? Are humans also better than artificial intelligence neural networks at identifying objects of unexpected sizes?
Researchers from the Department of Psychology and Brain Sciences at the University of Santa Barbara, in California, in collaboration with scientists from the Department of Computer Engineering at Ankara University, in Turkey, decided to study this issue. Scientists in California put 60 volunteers to the task of finding an object in a complex scene displayed on a computer screen. Each object was repeated three times in different scenes, but varying some characteristic, such as its color. After indicating on the screen the object they had to find with a word that defined it, the volunteers had a second to identify the object and its position.
When the objects to be found were displayed with a larger size, but inconsistent with the rest of the objects in the scene. For example, the rose bush was shown with the size of a tree.
The results showed that, despite the fact that the size of the objects that had to be found was greater and they occupied a greater percentage of the scene, the volunteers identified them much worse than when the objects were shown with the expected size.
On the other hand, when the scientists put three state-of-the-art neural networks (intelligent robots) to the same task, they did exactly as well or poorly when the object was displayed at the normal size as when it was displayed at the normal size. increased. It therefore appears that neural networks have also outperformed humans in this visual task.
However, scientists propose that the human brain has acquired the ability to learn the size of different objects and uses this information to locate them more quickly and rule out others. Obviously, in Nature and in daily life we do not find apples the size of watermelons, or remote controls the size of computer keyboards. If this conclusion is correct, as it seems, it is neural networks that still have a certain way to go to achieve the performance of our visual system in the everyday world, which is where they should be used.
More information on Jorge Laborda’s Blog: Intelligent vision and neural networks
References:
Eckstein et al., Humans, but Not Deep Neural Networks, Often Miss Giant Targets in Scenes, Current Biology (2017),
Popular works by Jorge Laborda
Kilo of Science Volume I. Jorge Laborda
Kilo of Science Volume II. Jorge Laborda
Kilo of Science Volume III. Jorge Laborda
Kilo of Science Volume IV. Jorge Laborda
Kilo of Science Volume V. Jorge Laborda
Kilo of Science Volume VI. Jorge Laborda
Kilo of Science Volume VII. Jorge Laborda
Kilo of Science Volume VIII. Jorge Laborda
Chained circumstances. Ed.Lulu
One moon, one civilization. Why the Moon tells us that we are alone in the Universe
One Moon one civilization why the Moon tells us we are alone in the universe
The intelligence funnel and other essays
#Cienciaes.com #Intelligent #vision #neural #networks