Yli-Krekola A bio-inspired computational model of covert attention and learning

The cerebral cortex of the mammalian brain is responsible for such things as processing sensory information, imagination, working memory and consciousness. The key functions of the cortex that help improve the behaviour of the animal can be thought to be learning regularities from the world, attending to behaviourally relevant targets and simulating a world model for the purpose of planning. This thesis builds a system-level computational model of learning and attention in the cortex. The model is inspired by the real cortex and it is intended to be later extended to include all the important functions in the cortex. The cortex learns a hierarchy of increasingly abstract representations of the world. The high levels can recognise complex objects and dynamical events and represent them with simple codes. Many models have been proposed for this kind of cortical learning. However , most of the models lack attention. The real cortex does not try to perceive everything simultaneously. Instead, it attends to those features in the world that are useful for the behaviour of the animal. Attention is also known to guide learning, which results in allocating the representational capacity in important features. There are computational models which have cortex-like attention, but so far, learning and attention have not been combined. In the model developed in this thesis, learning and attention support each other. Attention is not a distinct process which is controlled externally, but it emerges in the same network where learning and perception take place. Experiments with this model show that it is able to learn a hierarchy of increasingly invariant representations for objects and give birth to selective attention simultaneously. In the future, this model could be embodied in a controller of an autonomous robot.