What happens when the brain tries to create an image of the unknown? In a recently-published article, a Malmö University cognitive researcher describes why there is a limit to what humans can imagine but how that knowledge can be used within Artificial General Intelligence (AGI).

In the article Perception, Imagery, Memory and Consciousness, Magnus Johnsson, who is also a computer scientist, focuses on two perspectives: Firstly, a theoretical reasoning about how the biological architecture works in a mammalian brain. And secondly, how this knowledge can be implemented in an artificial cognitive architecture according to the same principles.

There is a limit even for normally developed brains.

Magnus Johnsson

“I assume that there is a high probability that there are a number of simple principles that natural cognition is built around. However, when these very simple principles are used on many different levels, the result is very complex abilities,” Johnsson explains, adding:

“Cognition researchers start from so-called "features", which are the brain's perception of an object. A feature can, for example, be the contrast separating a chair from its surroundings that the eye registers. When a chair is seen at a certain angle, the contrast will end up as a feature in a certain place in the cerebral cortex that depends on the angle. This is called topology-preserving (or order-preserving) representation. This architecture helps the brain to build an understanding of the visual impression.”

These topology-preserving maps are in turn divided into different sense modalities, depending on which sense registered the impression.

“We learn the simplest representations at a young age and they have a tendency to lock in. The higher up we get in the learning hierarchy, the more flexible and modifiable they are.”

Put simply, we can say that every time we see or imagine something, we activate the same representations to a great extent. And new things require new combinations of these. But there must be recognisable features at hand that have the potential to create the unknown. Or at least enough known representations, where we can then fill in the missing information ourselves.

As an example, Johnsson uses a famous cat experiment. Cats that were raised at an early age in constructed rooms without horizontal lines, later in life could not see lines as demarcated rooms, such as the meeting between floor and wall. It was too late for the cat to learn.

“There is a limit even for normally developed brains. An example is the spatial dimensions. We don't have the right representations to imagine four dimensions. Everything we can think of is limited by these feature building blocks. But such building blocks can theoretically be built in AI and used within AGI.”

“Then it is not enough with "deep learning", which is used in many areas today, and which provides deep knowledge in a narrow area. In order to build an artificial brain with such a capacity, it is necessary to create a complete cognitive system where all sensory modalities are connected to each other,” concludes Johnsson.

Text: Magnus Erlandsson & Adrian Grist