I started working in the Computational neuroscience group , led by Harri Valpola , in April 2006. I had been looking for a research group that would seriously try to model the brain using the present machine learning paradigms. This is to believe that the brain is not ultimately complicated, but implements a set of simple and powerful machine learning methods as a coordinated orchestra.
The vast amount of memory, familiar concepts and schemas, learned motor sequences etc. in the brain are due to the amount of similar neural circuits, not their complexity. When evolution invented some working circuits, it multiplied their number instead of developing new ones. This can be seen at least on the cerebral cortex and cerebellar cortex. Thus, reverse engineering the functions of evolutionarily new brain areas is a quite conceivable task. Evolutionarily old structures of the brain on the other hand might be more "hard coded" in the genes.
This group studies these functions and their interplay (system level neuroscience) by building a simple brain for an autonomous robot. It is supposed to be an autonomous controller, which can perform well in somewhat whole AI problems. Not just learning to perform specific control tasks, or learning to recognize objects, but to have internal motivations and capability to learn any kind of useful structure of the world. And knowing how to take advantage of these representations to satisfy its internal desires. Building the robot brain guides one to ask the right questions in neuroscience, and to neglect the insignificant details.
I'm building a neural network for the same purposes as what the
neocortex exists for. That is, learning a model of the world, which is
then used for object recognition, predictions, planning, etc. I'm
trying to find salient computational advantage for cortex-like
information processing, like:
- selective attention, which arises from local competition and
long-range excitation (biased-competition model)
- inference and learning that are heavily dependent on context
Analyzing a computational model of the cerebellum. And using this model in a more holistic system, which includes neocortex and basal ganglia like modules.
Spring 2010, assistant at the course Tfy-99.3730 Information processing in the brain.
Selective attention improves learning
In ICANN 2009, Part II, LNCS 5769, p. 285–294, 2009.
[article]
[poster]
Tietoisuuden evoluutio (the Evolution of Consciousness)
Talk at Tieteen päivät 2009 (Finnish popular science event, Days of Science )
Computational model of co-operating covert attention and learning
A. Yli-Krekola and H. Valpola.
Fifth Nordic Neuroinformatics Workshop, Espoo, Finland, p. 34, 2007.
[abstract]
[poster]
A bio-inspired computational model of covert attention and learning
Master's thesis, 2006. [pdf]