UC Irvine

In recent years, machine learning and neuroscience are increasingly intertwined. On the one hand, machine learning could benefit from the insights and inspiration provided by the discoveries in neuroscience, as well as the integration of biologically-inspired components. On the other hand, machine learning techniques can be used to enhance our knowledge of the brain and its functions. In this dissertation, we demonstrate how they could benefit from each other, with an emphasis on dynamic environments. We first introduce how machine learning could benefit from neuroscience. We start with two projects that integrate neuromodulatory systems into machine learning systems to handle dynamic environment changes. In the first project, we used a serotonergic (5-HT) neuromodulatory system to control the patience level of a mobile robot navigating in outdoor environments, resulting in flexible behaviors not typically found in traditional navigation solutions. The second project introduced a reinforcement learning solution augmented with noradrenergic (NE) and cholinergic (ACh) neuromodulation, enabling the agent to quickly adapt to dynamic environment changes. Besides the utilization of neuromodulatory systems, in the third project, we proposed a method of latent unified state representation (LUSR) to improve the domain adaptation performance of reinforcement learning methods by addressing the adaptation problem from the pixel domain to a latent space, inspired by the latent representation in brain. The fourth project introduced a method of policy distillation with selective input gradient regularization, inspired by memory consolidation, to achieve computation efficiency and high interpretability in explainable reinforcement learning. Finally, the dissertation discusses how machine learning can contribute to the field of neuroscience. The last project studied the transformation between the first-person view and global view, which utilized the machine learning technique of variational autoencoder (VAE) to enhance our understanding of how the brain conducts view transformation in a 3D environment. In summary, this dissertation demonstrated the mutually beneficial relationship between the fields of machine learning and neuroscience, highlighting how each field can help the other to achieve advancements in theory and practice.