Emergent embodied cognition in shallow, biological and artificial, neural networks

自然智能是在进化过程中为特定环境下的特定任务而形成的，并且能够适应终身快速而强大的学习。这些认知特性并不是为拥有大大脑的动物所保留的。蚂蚁和蜜蜂等昆虫在其自然觅食行为中表现出令人印象深刻的认知表现。例如，蜜蜂是导航专家，可以根据视觉记忆学习多种复杂的路线。同样地，蜜蜂也能很快地学会有回报的花朵的花型，从中采集花蜜和花粉。我们认为，昆虫在这些任务上令人印象深刻的表现源于专业化。也就是说，进化塑造了传感器、神经回路和行为，以解决花卉学习和导航任务。因此，尽管蜜蜂的大脑比人类小一百万倍，但它们在某些任务上的表现却比人类好。因此，理解感觉系统和学习策略如何适应任务和环境是至关重要的，比如小的神经回路产生复杂的认知。在这个项目中，我们将利用计算神经科学模拟速度的最新进展，系统地研究小型神经网络中认知的出现，并梳理出身体、大脑和环境的贡献。虽然这些问题与所有动物有关，但我们将重点关注昆虫。昆虫比脊椎动物更容易确定神经解剖学和观察神经生理学，并且我们对参与认知的神经回路（感觉叶、学习中心和运动控制区）有详细的描述。此外，昆虫表现出令人印象深刻的认知能力的特定行为在实验室和野外都得到了很好的描述。在这两个方面，我们对昆虫的理解比脊椎动物模型系统更详细和全面。昆虫的大脑不仅小而且浅。在感觉系统和运动输出之间只有几层处理。我们假设，这些浅层学习网络之所以如此有效，是因为它们与精心调整的感觉系统和行为相互作用。为了破译这些复杂的相互作用，我们将创建一个模拟世界和昆虫大脑关键区域的脉冲神经网络模型。利用这个模型，我们将研究受蜜蜂觅食启发的任务中的视觉认知特性。我们的类似蜜蜂的智能体将在3D模拟世界中移动，其感官系统可以复制昆虫的视觉，并作为蘑菇体（昆虫学习电路）的模型实现学习。至关重要的是，我们将独立操作这些组件，并使用优化方法来询问哪类感觉系统和行为产生最佳的学习性能。以前就设想过这种调查。然而，为了计算确定脑体和环境对认知的贡献，需要在agent的不同层次上对多个参数进行优化。针对这种多层次问题的优化方法（元学习或学习学习算法）是存在的，但这些方法的共同点是需要非常大量的评估。直到现在，我们才能够探索这些问题，部分原因是计算能力的提高，但也与我们最近在gpu上的峰值神经网络模拟速度和昆虫眼渲染技术方面的突破有关。因此，我们有机会在昆虫视觉认知的详细模型中研究感觉环境、大脑和行为之间的关系。对智能如何从浅层神经网络中产生的理解将是神经科学和认知科学的基础，但也有可能产生更自然的人工智能算法。

Natural intelligence has been shaped by evolution for specific tasks in specific environments and to be adaptable to allow for lifelong rapid and robust learning. These properties of cognition are not reserved for animals with large brains. Insects such as ants and bees show impressive cognitive performance within their natural foraging behaviours. For instance, bees are expert navigators and can learn multiple complex routes based on visual memories. Similarly, bees are able to rapidly learn the floral patterns of rewarding flowers from which to collect nectar and pollen. We believe that the impressive performance of insects on such tasks arises from specialisation. That is, evolution has shaped sensors, neural circuits and behaviours for solving flower learning and navigation tasks. Consequently, bees show better-than-human performance on some tasks despite having brains a million times smaller. It is thus crucially important to understand how sensory systems and learning strategies can be adapted to tasks and environments, such that small neural circuits produce complex cognition. In this project, we will leverage recent advances in the speed of computational neuroscience simulations to systematically investigate the emergence of cognition in small neural networks and to tease apart the contribution of body, brain and environment. While these issues are relevant to all animals, we will focus on insects. It is easier to determine neuroanatomy and observe neurophysiology in insects than vertebrates, and we have detailed descriptions of the neural circuits involved in cognition (sensory lobes, learning centres and motor control regions). Furthermore, the specific behaviours where insects demonstrate impressive cognition are very well described in the lab and in the wild. In both these regards, our understanding of insects is more detailed and comprehensive than for vertebrate model systems. Insect brains are not only small but also shallow. There are only a few layers of processing between sensory systems and motor output. We hypothesise that these shallow learning networks work so well because they are interacting with carefully tuned sensory systems and behaviours. In order to decipher these complex interactions we will create a simulated world and spiking neural network model of key insect brain regions. With this model, we will investigate properties of visual cognition in tasks inspired by the foraging of bees. Our bee-like agents will move in a 3D simulated world, with a sensory system which replicates insect vision and learning implemented as models of the Mushroom Bodies (the insect learning circuits). Crucially, we will independently manipulate these components and use optimisation methods to ask what classes of sensory system and behaviour produce the best learning performance.Investigations of this kind have been envisioned before. However, to computationally determine the contribution of brain body and environment to cognition necessitates optimisation of multiple parameters at different levels of the agent. Optimisation methods for such multi-level problems exist (meta-learning or learning to learn algorithms) but common to these approaches is the need for very large numbers of evaluations. It is only now that we can explore these questions, in part due to increases in computational power, but also in conjunction with our recent breakthroughs in the speed of spiking neural network simulations on GPUs and insect-eye rendering technology. We therefore have the opportunity to investigate the relationships between sensory environment, brains and behaviour in detailed models of insect visual cognition. The understanding of how intelligence emerges from shallow neural networks will be fundamental to neuroscience and cognitive science, but also has the potential to produce more natural AI algorithms.