FET: Small: Ferroelectric Transistor based Spiking Neural Networks with Adaptive Learning for Edge AI: from Devices to Algorithms

在电子系统中部署人工智能（AI）已被广泛认为是几个新兴应用的关键推动因素之一。然而，最先进的人工智能系统的能源效率和学习能力远远无法达到人类大脑的水平。这项研究进行了跨层探索，涵盖了新设备，低功耗神经网络和新的学习方案。该探索将利用具有内在神经模拟功能的铁电场效应晶体管（FeFET）技术，以实现节能的神经硬件和适应性学习。低功耗硬件解决方案和自适应学习算法有可能影响计算机辅助诊断、机器人、语音/人脸识别和数据分类等关键应用，从而直接使医疗保健、国防、安全等领域受益。此外，节能应转化为边缘设备的更长电池寿命和可穿戴健康监测平台等应用的节能数据处理。该项目将利用普渡大学的外展计划，并开发一个夏季本科生研究经验（REU）计划，让本科生和少数民族学生参与该项目。该项目的广泛性质将为本科生提供机会，以了解基于新兴技术的人工智能领域。尖峰神经网络（SNN），由于其自学习能力，在引入人工智能系统学习的适应性方面显示出希望，但精度较低。提高SNN的准确性和性能不仅需要支持适应性终身学习的新型学习机制，而且还需要一种适用于低功耗可扩展硬件的内在技术。为了满足这一关键需求，该项目将对基于SNN的多域FeFET进行全面的设备到算法探索。主要目标包括（a）使用FeFET设计低功率神经元和突触，以及（B）开发利用神经模拟设备的独特属性的自适应和顺序学习算法。为了实现FeFET的生物合理特征，将进行基于物理的器件优化，以利用铁电体的多畴效应和畴动力学。为了促进跨层探索，将开发一个设备到系统的仿真框架，捕获神经元和突触的丰富动态，它们在SNN中的相互作用，以及新学习算法对系统性能/准确性的影响。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

Incorporating artificial intelligence (AI) in electronic systems has been widely recognized as one of the key enablers for several emerging applications. However, the energy efficiency and the learning capabilities of state-of-the-art AI systems is far from that achievable by human brains. This research undertakes a cross-layer exploration spanning novel devices, low-power neural networks, and new learning schemes. The exploration will exploit Ferroelectric Field Effect Transistor (FeFET) technology with intrinsic neuro-mimetic features to achieve energy-efficient neural hardware and adaptable learning. The low-power hardware solutions and adaptive-learning algorithms have the potential to impact critical applications such as computer-aided diagnosis, robotics, speech/face recognition, and data classification, thereby directly benefiting areas such as healthcare, defense, security etc. Moreover, power savings should translate to longer battery life for edge devices and energy-efficient data processing for applications like wearable health-monitoring platforms. The project will leverage outreach programs at Purdue University and develop a summer Research Experiences for Undergraduates (REU) program to involve undergraduates and minority students in the project. The broad nature of this project will provide opportunity for undergraduate students to get introduced to the field of AI based on emerging technologies.Spiking Neural Networks (SNNs), due to their self-learning capabilities, show promise in introducing adaptability in learning for AI systems, but suffer from low accuracy. Improving SNN accuracy and performance not only necessitates novel learning mechanisms that support adaptable lifelong learning, but also an intrinsically suitable technology for low-power scalable hardware. To address this critical need, this project will carry out a comprehensive devices-to-algorithms exploration of multi-domain FeFET based SNNs. The main objectives include (a) design of low-power neurons and synapses using FeFETs, and (b) development of adaptive and sequential learning algorithms utilizing the unique attributes of the neuro-mimetic devices. To enable bio-plausible features in FeFETs, physics-based device optimization will be carried out to utilize the multi-domain effects and domain dynamics of ferroelectrics. To facilitate cross-layer exploration, a devices-to-systems simulation framework will be developed capturing the rich dynamics of the neurons and synapses, their interactions in an SNN, and the impact of new learning algorithms on system performance/accuracy.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.