Novel Neuromorphic Mechanisms and Structures

拟议的研究计划的目的是开发新的机械设备，使用类似于那些已经成功地用于人工神经网络在机器学习领域的概念。通过适应最初为软件中的机器学习开发的物理对象系统架构和设计方法，拟议的研究将导致全新的机械设备类别，这些设备可以实现复杂的功能，以自动化的方式设计简单，并且在尺寸和能耗方面非常高效。 在过去的几年里，我们已经为理解人工神经网络的一些最基本的特征做出了贡献，这些特征使其计算模型成为可能并导致其有利的特性，实际上可以在物理对象中实现，特别是在机械系统中。作为一个例子，我们已经表明，一个小的硅梁夹在两端的MEMS的非线性动力学可以用作能源效率，密集的神经形态计算的资源。我们还展示了3D打印超材料的第一个演示，其刚度是作用在超材料上的外力场模式的复杂函数，因此可以训练以高度特定的方式对外部负载做出响应。 拟议的工作包括直接在机械系统和结构中系统地研究机器学习概念的物理实现。我们已经证明，这一系列的研究可以产生功能原型，这代表了一种构建物理设备的新方法，为具有挑战性的应用提供解决方案。通过这项发现资助，机器学习领域的各种概念将被应用于机械对象，这些机械对象被设计为具有与人工神经网络中发现的某些属性相似的某些属性。因此，机械装置将具有以精细方式响应外部负载或刺激（加速度、声音）的能力。他们将接受训练，以获得这些复杂的反应，而不是被设计到最小的细节。它们有望继承神经网络卓越的泛化能力，对训练过程中从未见过的刺激做出充分的反应。拟议研究的主要预期成果将是支持新类别设备（MEMS，超材料等）的分析和设计方法。从长远来看，这些技术可以转移到工业领域，以更有效地解决高科技领域的问题，如患者健康监测、机器人控制、自动化制造、智能传感器和物联网。

The objective of the proposed research program is to develop novel mechanical devices using concepts similar to those which have so successfully been used with artificial neural networks in the field of machine learning. By adapting to physical objects system architectures and design methodologies that were initially developed for machine learning in software, the proposed research will lead to entirely new classes of mechanical devices which can implement complex functions, be simple to design in an automated manner, and be highly efficient in terms of size and energy consumption.     Over the last few years, we have contributed to the understanding that some of the most fundamental features of artificial neural networks, which enable their computing model and lead to their advantageous properties, can actually be realized in physical objects, and in particular in mechanical systems. As an example, we have shown that the non-linear dynamics of a small silicon beam clamped at both ends in a MEMS can be used as a resource for energy-efficient, dense neuromorphic computations. We have also presented the first demonstration of a 3D-printed metamaterial with a stiffness that is a complex function of patterns in the external force field acting on the metamaterial, and which can therefore be trained to respond in highly specific manners to external loads.     The proposed work consists in the systematic investigation of the physical implementation of machine learning concepts directly within mechanical systems and structures. We have already demonstrated that this line of research could yield functional prototypes which represent a new way of building physical devices, to provide solutions for challenging applications. With this Discovery grant, various concepts from the field of machine learning will be applied to mechanical objects that are designed to have certain properties that are similar to those found in artificial neural networks. As a result, the mechanical devices will have the ability to respond in elaborated ways to external loads or stimuli (acceleration, sound). They will be trained to acquire these complex responses, instead of being designed to the smallest detail. And they are expected to inherit the remarkable generalization capability of neural networks, to respond adequately to stimuli never seen during training. The main anticipated outcome of the proposed research will be an analysis and design methodology supporting new classes of devices (MEMS, metamaterials, etc.). In the long term, these could be transferred to the industry to more efficiently solve problems in high technology fields such as patient health monitoring, robot control, automated manufacturing, smart sensors and the Internet of Things.