FET: Medium: Neural network computation and learning in well-mixed and spatially-organized molecular systems

神经计算原理主要是在过去50年里发展起来的，它极大地提高了人们对大脑如何处理信息的理解，并成为现代机器学习技术的基础。直到最近，人们才开始意识到，神经计算的这些原理在多大程度上得到了更广泛的应用，并且可以在非神经系统中发现这些原理。特别是，神经计算原理在设计先进的合成分子系统方面的意义和潜力才刚刚开始被探索。关键原则包括使用线性阈值单元和赢者通吃竞争的高维模式识别的力量，从经验中学习以塑造网络连接强度，从网络动态共振中生成复杂模式，以及空间结构和计算能力之间的相互作用。使用DNA纳米技术的分子编程现在已经充分发展了方法来设计探索和利用这些神经计算原理的系统。这项研究将开创三种新型的DNA神经网络：能够在其环境中无监督学习复杂模式的DNA神经网络，能够作为反应扩散系统形成复杂空间模式的DNA神经网络，以及能够利用DNA纳米结构的分子尺度空间组织在某些操作制度下执行更有效的计算的DNA神经网络。从长远来看，神经计算原理与未来分子系统的结合将为从生物医学到材料科学的新应用打开大门。从中期来看，该项目的一个重要影响将是博士后和研究生的未来职业生涯，他们将从研究经验和指导中受益，无论他们是进入学术界，工业界还是企业家。在短期内，研究人员将把他们的科学理解融入在线软件工具，继续将研究与教育结合起来，为本科生提供有指导的暑期研究，增强跨学科的研究环境，并让更多的女性参与科学。他们还将通过出席公共活动、接受大众科学杂志采访和创作艺术品来说明他们的研究，从而增加公众对科学的参与。三种新型DNA神经网络开发的具体研究目标如下：首先，学习可以说是合成分子电路最理想的特性。该项目建立在研究人员目前的工作基础上，展示了监督学习，并将其扩展到更广泛的无监督学习类别。监督式学习的一个限制是，“老师”必须提供训练实例，说明应该学习什么。无监督学习解决了这一限制，它只让分子电路接触到它们遇到的东西，而不知道它们应该如何反应；这种新能力对于像细胞一样在分子环境中自主操作的分子机器人系统是必要的。其次，自从阿兰·图灵在形态发生方面的开创性工作以来，反应-扩散模式的形成一直被研究，因为它与生物模式有关，也是自组织的内在物理机制。然而，反应扩散模式的复杂性是有限的。研究人员将利用深度学习技术的最新突破，通过实例设计复杂的反应扩散网络。他们将使用可微编程方法，结合大型DNA神经网络合成的最新进展和可靠的DNA水凝胶作为反应扩散实验的空间基质。最后，研究人员将进行一项理论研究，应用他们在DNA折纸砖和表面局部化学反应网络方面的专业知识，为DNA神经网络引入一种新的计算架构。这种体系结构在设计复杂性和分子操作之间提供了一种新的权衡，随着网络规模的增加，它可能比以前的方法具有更好的可伸缩性。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The principles of neural computation, developed mostly over the past 50 years, have led to profoundly improved understanding of how brains process information and have served as the foundation for modern machine learning techniques. Only recently has it begun to be appreciated the extent to which these principles of neural computation apply more broadly and can be found operating within non-neural systems. In particular, the implications and potential of neural computing principles for designing advanced synthetic molecular systems are only beginning to be explored. Key principles include the power of high-dimensional pattern recognition using linear threshold units and winner-take-all competition, learning from experience to sculpt network connection strengths, the generation of complex patterns from the resonance of network dynamics, and the interplay between spatial structure and computing capabilities. Molecular programming using DNA nanotechnology now has sufficiently developed methods to design systems that explore and exploit these neural computing principles. This research will pioneer three new types of DNA neural networks: those capable of unsupervised learning of complex patterns in their environment, those capable of complex spatial pattern formation as reaction-diffusion systems, and those capable of exploiting the molecular-scale spatial organization of DNA nanostructures to perform more efficient computation in certain regimes of operation. In the long term, the incorporation of neural computation principles into future molecular systems will open doors to new applications ranging from biomedicine to materials science. In the medium term, an important impact of this project will be the future careers of the postdocs and graduate students who will benefit from the research experience and mentoring, whether they move on to academia, industry, or entrepreneurship. In the short term, the researchers will incorporate their scientific understanding into online software tools, continue to integrate research with education, provide undergraduates with mentored summer research, enhance the interdisciplinary research environment, and involve more women in science. They will also increase public engagement with science by presenting at public events, interviewing for popular science magazines, and creating artworks to illustrate their research.Specific research goals for the development of the three new types of DNA neural networks are as follows. First, learning is arguably the most desirable property of synthetic molecular circuits. This project builds on the researchers’ current work demonstrating supervised learning and expands it to the broader category of unsupervised learning. A limitation of supervised learning is that a “teacher” must provide training examples that indicate what should be learned. Unsupervised learning addresses this limitation by exposing the molecular circuits to only what they encounter but not how they should respond; this new capability would be necessary for molecular robotic systems that operate autonomously, as cells do, within a molecular milieu. Second, reaction-diffusion pattern formation has been studied since Alan Turing’s seminal work on morphogenesis, both for its relevance to biological patterning and as an intrinsic physical mechanism of self-organization. However, the complexity of reaction-diffusion patterns has been limited. The researchers will leverage recent breakthroughs in deep learning techniques to design complex reaction-diffusion networks by example. They will use the differentiable programming approach, combined with the recent advances in the synthesis of large DNA neural networks and reliable DNA hydrogels as a spatial substrate for reaction-diffusion experiments. Finally, the researchers will perform a theoretical study that applies their expertise in DNA origami tiles and surface-localized chemical reaction networks to introduce a novel computing architecture for DNA neural networks. This architecture provides a new trade-off between design complexity and molecular operation that may scale better than prior approaches as the network size increases.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.