| Learning in the brain is possible because the connection strengths between neurons are modifiable. For learning stimulus-response mappings, there is at present no theory that provides a strategy for modifying connections strengths that is both powerful and biologically plausible. Supervised learning is powerful but biologically implausible. Reinforcement learning is biologically plausible, since learning is only guided by rewards and punishments, but it is comparatively inefficient. It lacks a mechanism that can identify units at early processing levels that play a decisive role in the network?s input-output mappings. For categorization tasks, our previous work demonstrated that this so-called credit-assignment problem can be solved by a new role of feedback connections. In our learning scheme, two factors determine plasticity of connections: (1) a reinforcement signal that is homogeneous across the network and depends on the amount of reward obtained after a learning trial, and (2) a feedback signal from the output layer that limits plasticity to those units at earlier processing levels that are crucial for the input-output mapping. The influence of this feedback signal corresponds to the effects of selective attention in the brain. We therefore call our new scheme attention-gated reinforcement learning (AGREL). The present proposal investigates whether AGREL can serve as a general learning algorithm in complex input-output mappings, regression tasks, and delayed and sequential decision tasks. It is likely that we will develop new algorithms that accomplish these tasks, and that AGREL will become a learning algorithm that is as powerful as supervised learning and yet biologically plausible. |
