Computational Neuroscience

Research Group at Laboratory of Computational Engineering

 

The computational neuroscience group in LCE studies the system-level organization of the brain. In the brain, there are several interacting subsystems which work in concert and each contribute to generation and adaptation of behavior. In order to understand the brain, our group is building computational models of these subsystems and studying the complex, emergent behavior and learning of an agent which interacts with its environment. This type of research is called embedded computational neuroscience and it requires a body and environment to interact with. To this end, we have used Webots simulator platform. In other words, we have worked on simulated robots but in the future we also aim at verifying the results with real robotic platforms.

During evolution, brain was always part of a complete autonomous system. We are roughly following the evolutionary path of mammalian brain development and that is why our first embodied model - a complete brain which controls an autonomous robot - was cerebellar motor control. Cerebellum is an evolutionarily old system and is shared by all vertebrates and is critically involved in motor control. However, it is also known that cerebellum is involved in sensory processing, sensorimotor integration and cognitive functions. Algorithmically, a simple description of cerebellum is that it is a predictor. We have shown how a simple predictor can assist in motor control and it is also easy to see how a predictor can assist in the other tasks in which cerebellum is known to be involved.

Our development of the computational models for different parts of the brain is by no means strictly serial. Rather, we are developing and testing computational models of various parts of the brain before they get integrated in autonomous agents. The point in building a complete autonomous agent is that we get a better intuition about what kind of processing is needed by the already existing integrated components in order to improve motor performance and ability to tackle more complex environments. We can then tune the hypotheses about the computational role of different parts of the mammalian brain.

Apart from the cerebellar model, the main topic of our research has so far been the mammalian neocortex. As opposed to cerebellum, neocortex has appeared quite late in evolution and is shared by modern mammals. Neocortex is the most complex part of the brain and tremendously enlarged in humans. It is the site of high-level cognition, consciousness and imagination. Dorsal cortex, the evolutionary precursor of neocortex is much older and simpler, though. It seems that one of the first tasks the precursor of neocortex solved was development of behaviourally meaningful represenations. As an example, consider a balancing robot (described in more detail in the next section) which is riding an an uneven terrain and has cameras enabling depth vision - in principle; it is by no means a trivial task to extract depth information from the images of two cameras. We are investigating how motor signals can bias the development of perceptual systems and dynamic selection of useful information (selective attention) such that perception will be optimized for the behavioural needs of motor control.

Other systems that we plan to incorporate in the model later include basal ganglia and hippocampal formation. Basal ganglia are thought to assist in selection of behaviours and learning through trial and error. Hippocampal formation appears to have develop to assist navigation and is able to compress, store and replay sequences of events.

Although the different modules of the mammalian brain seem to have evolved to serve the needs of basic motor control, these mechanisms have later been adopted by higher-level cognition. For instance, when we plan, we in a sense navigate through options and select promising paths of thinking. This can be considered as internalised navigation where basal ganglia help us make choices and hippocampal formation enables us to remember the paths of thinking. It is easier to study and understand navigation, manipulation and sensory associations than planning, reasoning and symbolic language, but the same underlying mechanisms are at work.

Computational Neuroscience Research Group
Laboratory of Computational Engineering
Helsinki University of Technology
P.O.Box 9203, FI-02015 TKK, Finland