CAREER: SHF: Bio-Inspired Microsystems for Energy-Efficient Real-Time Sensing, Decision, and Adaptation

当代人工智能系统通常不能实时适应其环境，非常耗电，并且通常是独立于底层硬件开发的。然而，生物智能已经解决了这些限制，具有能效、适应性和与底层基质的良好集成。该项目以生物智能为指导原则，开发了微电子系统，这些系统可以高效地感知来自环境的物理信号，做出实时决策，并以最小的能源使用来适应和学习。通过从生物学中汲取灵感，这个项目通过将记忆、计算、感知和学习算法无缝地结合在一起，模糊了感知、计算和算法之间的区别。通过模仿这种生物模型，该项目寻求开发下一代自主智能系统，应用范围从更智能的手机到改进的脑机接口。概述的教育活动将使与行业合作伙伴和历史上一直是黑人的学院和大学能够开展尖端微电子教育，使国内劳动力能够满足国家的战略半导体需求。该项目共同设计持续学习算法，采用尖端、节能的微电子设计，利用铁电场效应晶体管(FeFET)形式的新兴设备，实现下一代、高能效、自适应的硬件，用于传感、决策和学习。利用FeFET作为可编程跨导的模拟到特征转换器前端系统将被设计为获取模拟输入并提取相关的学习特征。这些子系统将为下游基于FeFET的内存计算(CIM)电路提供定制设计的电路，以缓解目前限制大多数CIM架构的模数转换器瓶颈。静态随机存取存储器将增强FeFET结构，以实现片上学习和动态可重新配置。与底层硬件同步，定制的连续学习算法将与模数转换器和FeFET阵列共同设计，以赋予系统高能效的弹性和自适应能力。使用该方法设计的微电子系统可广泛应用于从脑机接口和植入型系统到无线系统的盲波形分类。为了验证该方法的有效性，我们将制作一个集成电路并进行测试。这些还将为进一步完善和校准性能评价和设计空间探索的软件模型提供数据。该项目最终将开发对生物灵感、能源效率和自主代理至关重要的组件。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Contemporary artificial intelligence systems typically do not adapt to their environment in real time, are very power hungry, and are typically developed in isolation from the underlying hardware. However, biological intelligence has addressed these limitations, being energy-efficient, adaptable, and well-integrated with the underlying substrate. Using biological intelligence as the guiding principle, this project develops microelectronic systems that can efficiently, sense physical signals from their environment, make real-time decisions, and adapt and learn with minimal energy usage. By drawing inspiration from biology, this project blurs the distinction between sensing, computation, and algorithm by seamless integration of memory, computing, and sensing and a learning algorithm. Through mirroring this biological model, the project seeks to develop the next generation autonomous intelligent systems with applications ranging from smarter cellphones to improved brain-machine interfaces. The outlined educational activities will enable collaboration with industrial partners and historically black colleges and universities to enable cutting edge microelectronic education to enable the domestic workforce to meet the strategic semiconductor needs of the Nation.The project co-designs continual learning algorithms with cutting-edge, energy-efficient, microelectronic designs that leverage emerging devices in the form of Ferroelectric Field-Effect Transistors (FeFETs) to enable next-generation, energy-efficient, adaptive hardware for sensing, decision-making, and learning. Analog-to-Feature converter front-end systems leveraging FeFETs as programmable transconductances will be designed to acquire analog input and extract pertinent learned features. These subsystems will feed downstream FeFET-based compute-in-memory (CIM) circuits with custom-designed circuits to alleviate the analog-to-digital converter bottleneck currently limiting most CIM architectures. Static Random Access Memory will augment FeFET structures to enable on-chip learning and dynamic reconfigurability. In lockstep with the underlying hardware, tailored continual learning algorithms will be co-designed with the analog-to-digital converters and the FeFET array to endow the system with energy-efficient resilience and adaptation. Microelectronic systems designed using the presented approach could see wide ranging applications from brain-computer-interfaces and implantable systems to blind waveform classification for wireless systems. To validate the approach, an integrated circuit will be fabricated and measured. These will also serve to provide data to further refine and calibrate software models for performance evaluation and design-space exploration. This project will ultimately develop components critical for biologically inspired, energy-efficient, autonomous agents.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.