ActiveAI - active learning and selective attention for robust, transparent and efficient AI

我们将把昆虫生物学和神经科学领域的世界领导者与生物机器人建模和计算神经科学领域的世界领导者聚集在一起，建立一种合作伙伴关系，这将在理解昆虫、机器人和人工智能（AI）自主系统的主动学习和选择性注意方面具有变革性。通过考虑大脑，行为和环境如何在自然动物行为中相互作用，我们将开发新的算法和方法，用于动态现实世界应用的自主机器人和人工智能的快速，稳健和有效的学习。人工智能的最新进展，尤其是深度学习，已经被证明在为特定复杂问题创造解决方案方面取得了令人难以置信的成功（例如，在围棋中击败最好的人类棋手，在城市中驾驶汽车）。但随着我们对这些方法了解的越来越多，它们的局限性也变得越来越明显。例如，深度学习解决方案通常需要大量的计算能力，极长的训练时间和非常大量的标记训练数据，而这些数据根本无法用于许多任务。虽然他们非常擅长解决特定的任务，但在将这些知识转移到其他密切相关的任务上，他们可能相当糟糕（而且不可预测）。最后，科学家和工程师正在努力理解他们的深度学习系统学到了什么，以及他们学得有多好。与昆虫自然进化出来的智力相比，这些局限性尤其明显。昆虫当然不会下围棋或开车，但它们在进化过程中表现得非常出色。例如，与目前的任何人工智能系统不同，蚂蚁学会如何利用它们微小的大脑提供的有限计算能力和对世界的最小探索来有效地觅食。我们认为这种差异的产生是因为自然智能是大脑-身体-环境相互作用的闭环特性。进化的先天行为与专门的传感器和神经回路协调一致，以最大的效率提取和编码与任务相关的信息，并辅以选择性注意机制，将学习集中在与任务相关的特征上。这种对行为具身代理的关注在目前的人工智能技术中代表性不足，但为上述问题提供了解决方案，这可以通过对人工智能的原始定义进行研究来实现：对生物学习和智能的描述和模拟，既复制了动物的能力，又揭示了智能的生物学基础。这项工作需要研究行为动物的大脑运作，因为了解神经活动如何与环境、身体和行为相互作用，并受其影响，以及与选择性注意的相互作用，是至关重要的。这些实验现在可以通过结合苍蝇和食蚜蝇的神经记录（可以识别选择性注意的神经标记）的最新进展，再结合蚂蚁的虚拟现实实验；澳大利亚团队首创的技术结合验证现实世界中机器人平台上大规模神经模型的新兴假设，这是英国团队首创的一种方法，该项目代表了一个独特而及时的机会，可以改变我们对动物学习的理解，并通过这种方式改变机器人和人工智能系统的学习。我们将创造一个跨学科的合作研究环境，实现实验、分析、计算和机器人建模的“良性循环”。新的发现在这个良性循环中前前后后，每一个学科都为其他学科提供信息，以产生对主动学习和选择性注意如何使小脑昆虫学习复杂世界的功能性理解。通过这种理解，我们将开发ActiveAI算法，该算法在学习和最终网络配置方面效率高，对现实世界条件具有鲁棒性，并且学习速度快。

We will bring together world leaders in insect biology and neuroscience with world leaders in biorobotic modelling and computational neuroscience to create a partnership that will be transformative in understanding active learning and selective attention in insects, robots and autonomous systems in artificial intelligence (AI). By considering how brains, behaviours and the environment interact during natural animal behaviour, we will develop new algorithms and methods for rapid, robust and efficient learning for autonomous robotics and AI for dynamic real world applications.Recent advances in AI and notably in deep learning, have proven incredibly successful in creating solutions to specific complex problems (e.g. beating the best human players at Go, and driving cars through cities). But as we learn more about these approaches, their limitations are becoming more apparent. For instance, deep learning solutions typically need a great deal of computing power, extremely long training times and very large amounts of labeled training data which are simply not available for many tasks. While they are very good at solving specific tasks, they can be quite poor (and unpredictably so) at transferring this knowledge to other, closely related tasks. Finally, scientists and engineers are struggling to understand what their deep learning systems have learned and how well they have learned it. These limitations are particularly apparent when contrasted to the naturally evolved intelligence of insects. Insects certainly cannot play Go or drive cars, but they are incredibly good at doing what they have evolved to do. For instance, unlike any current AI system, ants learn how to forage effectively with limited computing power provided by their tiny brains and minimal exploration of their world. We argue this difference comes about because natural intelligence is a property of closed loop brain-body-environment interactions. Evolved innate behaviours in concert with specialised sensors and neural circuits extract and encode task-relevant information with maximal efficiency, aided by mechanisms of selective attention that focus learning on task-relevant features. This focus on behaving embodied agents is under-represented in present AI technology but offers solutions to the issues raised above, which can be realised by pursuing research in AI in its original definition: a description and emulation of biological learning and intelligence that both replicates animals' capabilities and sheds light on the biological basis of intelligence.This endeavour entails studying the workings of the brain in behaving animals as it is crucial to know how neural activity interacts with, and is shaped by, environment, body and behaviour and the interplay with selective attention. These experiments are now possible by combining recent advances in neural recordings of flies and hoverflies which can identify neural markers of selective attention, in combination with virtual reality experiments for ants; techniques pioneered by the Australian team. In combination with verification of emerging hypotheses on large-scale neural models on-board robotic platforms in the real world, an approach pioneered by the UK team, this project represents a unique and timely opportunity to transform our understanding of learning in animals and through this, learning in robots and AI systems. We will create an interdisciplinary collaborative research environment with a "virtuous cycle" of experiments, analysis and computational and robotic modelling. New findings feed forward and back around this virtuous cycle, each discipline informing the others to yield a functional understanding of how active learning and selective attention enable small-brained insects to learn a complex world. Through this understanding, we will develop ActiveAI algorithms which are efficient in learning and final network configuration, robust to real-world conditions and learn rapidly.