Neuromodulated networks for lifelong learning and adaptation

The development of robust and adaptable intelligent system has been a long standing grand challenge. Recently, machine learning methods via neural networks have gained prominence. However, most systems are excellent at solving a single task. The field of lifelong (machine) learning seeks to alleviate this problem by endowing intelligent agents with the ability to learn multiple tasks. Interesting questions arise from the field about how to design agents that can learn many tasks over a lifetime without forgetting, rapidly adapt to task changes, reuse existing knowledge to foster rapid learning of new tasks, use knowledge from task being learned to improve knowledge of previous tasks. This thesis investigates several of the aforementioned challenges through the use of biologically inspired neuromodulatory mechanisms, that are incorporated into standard artificial neural networks. Experiments were conducted in simulated reinforcement learning environments in domains such as navigation, robotics, and autonomous driving simulations. A common theme from the findings showed that neuromodulation enabled the lifelong learning systems to solve problems of increasing complexity in comparison to systems without neuromodulation. Also, neuromodulation enabled the rapid switch of learned behaviour via the dynamic regulation of the agent’s neural activity in fully and partially observable scenarios, and the efficient learning of new tasks in an online manner without forgetting. However, the use of neuromodulation incurs an extra cost in computational and memory requirements in neural networks.