详情

原文传递 Adaptive Learning Through Active Neuromodulation (ALAN).

题名：	Adaptive Learning Through Active Neuromodulation (ALAN).
作者：	Aguilar-Simon, M; Brna, A; Brown, R; Folsom, L; Cook, J; Park, S; Yanoschak, A; Shimizu, R.
摘要：	Under the lifelong learning machines (L2M) program, Teledyne set out to research, implement and demonstrate algorithmic approaches toaddress two key problems. First, enabling intelligent agents to self-supervise in order to adapt and learn in complex environments withoutexternal intervention. To address this problem, Teledyne developed and validated the role of uncertainty tracking and modulation to enableagents to monitor their own performance and adapt with confidence when appropriate conditions exist. This is a significant breakthrough in thatit demonstrated self-supervised learning and task performance by an embodied agent. The second problem was to enable robust knowledgerepresentations that would remain accurate despite continual learning and adaptation and could accommodate the complexity of learningmultiple tasks with likely differing requirements on knowledge granularity and composition. Teledyne developed and implemented ahierarchical learning system capable of decomposing task information across multiple layers to maximize robustness and re-use. This is asignificant breakthrough as it enables a new class of learning systems that can maintain a consistent knowledge base and update it toaccommodate multiple tasks without the requirements that they share a unified representation. The resulting algorithms were demonstrated tohave a role in enhancing performance in state-of-the-art machine learning systems, and thus can be incorporated into many present-day AIsolutions to equip them with lifelong capabilities. A key recommendation is to look for opportunities to transition these capabilities intoexisting AI systems, thus facilitating their transition to next-wave AI. Another recommendation is to see these accomplishments as a first step inelucidating lifelong learning mechanisms, and to engage in continued research to more fully understand how to achieve learning in highlycomplex environments and conditions. These will likely force us to look more closely at a more complete solution for building, maintaining andleveraging hierarchical knowledge representations.
总页数：	84 pages