Context dependent and multimodal learning: from insect brains to robot controllers

当前机器人技术的一个核心问题是如何扩展到更复杂的认知能力，例如上下文相关的学习和预测。虽然昆虫通常被视为简单的反射系统，但它们实际上比任何现有的自主机器人都更有能力。他们能够学习，整合多感官线索，现实世界的导航和灵活的行为选择。由于它们以相对较小的大脑获得了这种能力，因此了解这些机制应该会带来高效的机器人应用。在生物学中，最近对昆虫大脑的理解有了很大的兴趣和实质性的进展。到目前为止，这些系统的建模已经落后，但由于许多原因，它是必不可少的。通过构建昆虫大脑的模型，我们可以精确地评估关于其功能的假设，并测试哪些元素对复杂行为至关重要。此外，通过在机器人硬件中实现这些假设，我们可以在真实的行为环境中理解系统。因此，有一个真实的机会，有助于生物学知识的同时，开发系统，有有用的应用程序作为机器人控制器。我们在这个项目中的目的是开发和评估昆虫大脑的学习模型，使用生物实验，计算建模和硬件实现的组合。特别是，我们想研究支持比简单联想更复杂的学习形式的神经机制。这些包括上下文依赖，概括，和基于预期的反应表达。深入了解这些能力需要更密切地关注昆虫机制的细节。例如，了解学习的不同阶段和时间尺度以及这些如何由不同的生物化学过程支持可能很重要。我们可以利用各种情况的理想组合，在这一领域取得实质性进展。PI（Webb）在构建昆虫行为机器人模型方面拥有丰富的经验，包括在硬件中实现感官和神经处理机制。沿着研究人员-CI（Wessnitzer），她开发了相关昆虫大脑机制的初始模型，并与该领域的领先生物学家有着密切的联系。其中之一是CI（阿姆斯特朗），他正在使用先进的遗传技术来确定昆虫大脑中不同结构和信号通路的作用。因此，我们打算开发一个紧密相连的模式，其中：-行为实验模型提供数据模型评估; -硬件实现的模型提供了真实的世界的评价和激励抽象; -抽象的模型建议的关键功能的角色，可以使用遗传操作的昆虫进行测试。其结果将大大提高对昆虫大脑的理解，并朝着机器人认知控制器迈出实质性的一步。

A central issue in current robotics is how to scale up to more complex cognitive abilities, such as context dependent learning and prediction. Although insects are often viewed as simple reflexive systems, they are in fact more competent than any existing autonomous robots. They are capable of learning, integration of multisensory cues, real-world navigation, and flexible behavioural choice. As they obtain such competences with relatively small brains, understanding these mechanisms should lead to efficient robot applications. Within biology, there has recently been great interest and substantial advance in understanding the insect brain. So far, modelling of these systems has lagged behind, but it is essential for many reasons. By building models of the insect brain we can evaluate precisely expressed hypotheses about its function, and test which elements are crucial for complex behaviour. Moreover by implementing these hypotheses in hardware on robots we can understand the systems in real behavioural contexts. Thus there is a real opportunity to contribute to biological knowledge at the same time as developing systems that have useful application as robot controllers. Our intention in this project is to develop and evaluate models of learning in insect brains, using a combination of biological experiments, computational modelling, and hardware implementations. In particular we want to examine the neural mechanisms that support forms of learning more complex than simple association. These include context dependence, generalisation, and expectation-based expression of responses. Insight into these capabilities requires closer attention to the details of the mechanisms in the insect. For example, it may be important to understand the different stages and time-scales of learning and how these are supported by different biochemical processes. We can exploit an ideal combination of circumstances to make substantial advances in this area. The PI (Webb) has extensive experience in building robot models of insect behaviour, including implementing sensory and neural processing mechanisms in hardware. Along with the researcher-CI (Wessnitzer) she has developed initial models of the relevant insect brain mechanisms, and has strong connections to the leading biologists working in this area. One of these is the CI (Armstrong) who is using advanced genetic techniques to determine the roles of different structures and signalling pathways in the insect brain. Thus we intend to develop a tightly linked paradigm in which: - behavioural experiments suggested by the models provide data for model evaluation; - hardware implementation of the models provides real world evaluation and motivates abstraction; - abstracted models suggest key functional roles that can be tested using genetic manipulations on the insects. The outcome will be both significantly improved understanding of insect brains and a substantial step towards cognitive controllers in robots.