EFRI BRAID: Emulating Cerebellar Temporally Coherent Signaling for Ultraefficient Emergent Prediction

EFRI BRAID：模拟小脑时间相干信号以实现超高效紧急预测

• 批准号: 2317974
• 负责人: Mark Hersam
• 金额: $ 200万
• 依托单位: Northwestern University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-09-01 至 2027-08-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2317974&HistoricalAwards=false
• 关键词: EFRI; BRAID; Emulating; Cerebellar; Temporally

Although artificial intelligence (AI) has been applied to many computational problems, biological intelligence remains superior to AI for most cognitive tasks. For example, the brain is constantly receiving and ignoring massive volumes of sensory signals yet remains perpetually vigilant to anomalies in order to respond rapidly to unanticipated inputs. In contrast, modern AI performs poorly on similar tasks, requiring extensive training and propagation through multi-layer artificial neural networks. Consequently, robust anomaly detection in AI is slow and energy inefficient, posing challenges for high-value applications such as cybersecurity. Neuroscience research has shown that the cerebellum allows anomaly detection to emerge through contextual prediction, pattern separation, and response actuation. In an effort to emulate cerebellar functions, this project develops electronic devices that switch between asynchronous and synchronous behavior when triggered by sensory inputs. These devices are derived from nanoelectronics materials that realize temporally coherent signaling for diverse applications including cybersecurity, autonomous robotics, and power-delivery control. In addition, this project comprehensively analyzes the ethical, legal, and societal implications of the proposed research in collaboration with multiple stakeholders including college students, educators, and community workers. To ensure that these transformative outcomes are communicated to the most diverse audiences, multiple education and outreach initiatives aim to broaden participation among underrepresented and marginalized sections of society.Neuromorphic hardware chips are emerging as disruptive technologies to process and categorize vast amounts of digital data. The majority of the current implementations are based on well-studied feed-forward and recurrent neuronal architectures of the mammalian cerebrum and are thus optimized to perform only certain types of classification tasks. In contrast, theoretical neuroscience concepts derived from the cerebellum are underrepresented in artificial intelligence hardware even though the cerebellum has evolved to efficiently solve a wide range of problems such as anomaly detection in complex and noisy environments. Cerebellar accuracy and robustness are achieved by a unique neuronal coding architecture based on high firing rates, temporally coherent signaling, and complex spiking. To achieve similar functionality, this project develops electronic hardware that emulates the essential features of cerebellar neuronal coding. The resulting bio-realistic implementations are tested against the use cases of anomaly detection in cybersecurity, autonomous robotics, and power-delivery control. Specifically, this cross-disciplinary project combines ideas from theoretical neuroscience, materials science, and computer engineering to develop hardware prototypes based on two-dimensional semiconductors and van der Waals heterojunctions including synaptic devices based on memtransistors and spiking neurons based on Gaussian heterojunction transistors. In addition, this project comprehensively analyzes the ethical, legal, and societal implications of the proposed research in collaboration with multiple stakeholders including college students, educators, and community workers. To ensure that these transformative outcomes are communicated to the most diverse audiences, multiple education and outreach initiatives aim to broaden participation among underrepresented and marginalized sections of society.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.

尽管人工智能（AI）已经应用于许多计算问题，但在大多数认知任务中，生物智能仍然优于人工智能。例如，大脑不断地接收和忽略大量的感官信号，但为了对未预料到的输入迅速作出反应，它始终对异常保持警惕。相比之下，现代人工智能在类似任务上表现不佳，需要通过多层人工神经网络进行大量训练和传播。因此，人工智能中的鲁棒异常检测速度缓慢且能源效率低下，给网络安全等高价值应用带来了挑战。神经科学研究表明，小脑可以通过情境预测、模式分离和响应驱动来进行异常检测。为了模拟小脑的功能，这个项目开发了一种电子设备，当被感官输入触发时，它可以在异步和同步行为之间切换。这些器件由纳米电子材料制成，可实现各种应用的暂时相干信号，包括网络安全、自主机器人和电力输送控制。此外，该项目还与包括大学生、教育工作者和社区工作者在内的多个利益相关者合作，全面分析拟议研究的伦理、法律和社会影响。为确保将这些变革性成果传达给最多样化的受众，多项教育和外联举措旨在扩大社会中代表性不足和边缘化群体的参与。神经形态硬件芯片正在成为处理和分类大量数字数据的颠覆性技术。目前的大多数实现都是基于对哺乳动物大脑的前馈和循环神经元结构的充分研究，因此仅针对某些类型的分类任务进行了优化。相比之下，尽管小脑已经进化到能够有效地解决复杂和嘈杂环境中的异常检测等广泛问题，但来自小脑的理论神经科学概念在人工智能硬件中的代表性不足。小脑的准确性和鲁棒性是通过基于高放电率、时间相干信号和复杂尖峰的独特神经元编码结构实现的。为了实现类似的功能，该项目开发了模拟小脑神经元编码基本特征的电子硬件。通过对网络安全、自主机器人和电力输送控制中的异常检测用例进行测试，得出了逼真的生物实现。具体来说，这个跨学科项目结合了理论神经科学、材料科学和计算机工程的思想，开发基于二维半导体和范德华异质结的硬件原型，包括基于mem晶体管的突触器件和基于高斯异质结晶体管的尖峰神经元。此外，该项目还与包括大学生、教育工作者和社区工作者在内的多个利益相关者合作，全面分析拟议研究的伦理、法律和社会影响。为确保将这些变革性成果传达给最多样化的受众，多项教育和外联举措旨在扩大社会中代表性不足和边缘化群体的参与。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。