Skip to Main Content
A significant feature of brain intelligence is flexibility. This is generally lacking in current machine intelligence We think that learning that effectively uses the combination of multiple information representations is the key to constructing flexible machine intelligence. This hypothesis is demonstrated by means of a simple connectionist model of intrinsically motivated reinforcement learning. A linear approximation of reward functions that depends on multiple representations is engaged in our model. We show preliminary results for a model network that enables a flexible learning response to several different situations. Multiple representations in our model accelerate the learning not only in complex situations that need many kinds of information, but also in simple situations.