Mail to:
cnb@atr.jp
@
METALEARNING, NEUROMODULATION AND EMOTION
Research Supported by "Creating
the Brain",
CREST, Japan
Science and Technology Corporation
The aim of this research project is to elucidate the higher-level mechanisms for coordinating learning modules in the brain. We develop computational theories of metalearning in adaptive systems and investigate how they could be implemented in the brain by combining animal, human, and robotic experiments. This would allow us to better understand the functions of emotional systems in regulating human behaviors, to understand the causes of mental and behavioral disorders, and to build robots and communication agents that are truly adaptive, autonomous, and more human-like.
Recent
advances in artificial neural networks and machine learning have allowed
us to build robots and virtual agents that can learn a variety of goal-directed
behaviors. However, the performance of learning systems depends critically
on a number of metaparameters that controls how the detailed system
parameters change with learning. The optimal settings of metaparameters
are tasks and environment dependent and in most learning systems today,
they are hand-tuned by human experts.
The fact that humans and animals can learn new behaviors under unknown environments suggests that the brain has a mechanism for actively tuning metaparameters, such as the speed of learning, the size of noise for exploration, and the time scale of evaluation. The possible substrate for such metalearning is a number of neuromodulators, such as acetylcholine, noradrenaline, serotonin, and dopamine, that project diffusively to the entire brain.
Based on the theory of reinforcement learning and data from neurobiological experiments, we propose a coherent hypothesis about how these neuromodulators are used for regulating metaparameters for learning and test it by way of animal experiments, human patient studies, computer simulations, and robotic experiments.
@
@