SHF: Small: Energy and Computational Efficient Deep Generative AI Models via Emerging Devices, Circuits, and Architectures

深度生成人工智能（AI）模型可以学习复制其输入或其输入的变化版本。 然而，需要解决的一个关键挑战是它们的能源和计算成本。通过新兴设备，电路和架构中的新方法，对深度生成AI硬件和软件的设计，验证，操作和评估进行基础研究是可取的。人工智能的能量和计算成本已经成为其在真实的世界中应用的瓶颈。研究和教育将通过课程和实验室开发进行整合。该项目将通过西班牙裔专业工程师协会（Society of Hispanic Professional Engineers）、美国黑人工程师协会（National Society of Black Engineers）和女性工程师协会（Society of Woman Engineers）招募代表性不足的学生和女性学生。该项目的目标是开发新的生成式人工智能模型，其设计和架构比目前更简单。通过从设备和电路到架构和算法的努力，探索了设计深度学习硬件加速器的新途径。研究了基于卷积的网络中关键操作的基于细胞神经网络的实现，因为它允许在模拟域中执行与深度生成AI模型相关的大部分计算。 研究了混合信号电路和架构的发展，这些电路和架构通过利用新兴器件技术的独特物理特性来实现最佳的深度生成网络设计。该项目预计将为深度生成AI模型带来能量和延迟的数量级改善，这将促进其应用并使AI行业受益。该奖项反映了NSF的法定使命，并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，被认为值得支持。

Deep generative artificial intelligence (AI) models can learn to reproduce their inputs or the variational versions of their inputs. However, a critical challenge that needs to be addressed is their energy and computational cost. Foundational research in the design, verification, operation, and evaluation of deep generative AI hardware and software through novel approaches in emerging devices, circuits, and architectures is desirable. The energy and computational cost of AI has become a bottleneck for its applications in the real world. Research and education will be integrated through course and lab development. Under-represented and women students will be recruited for this project through the Society of Hispanic Professional Engineers, National Society of Black Engineers, and Society of Woman Engineers.This project targets the development of new generative AI models with simpler designs and architecture than are currently available. A novel path is explored for designing deep-learning hardware accelerators via efforts that span from devices and circuits to architectures and algorithms. The Cellular Neural Network-based realizations for key operations in convolution-based networks is studied, because it allows the bulk of the computation associated with a deep generative AI model to be performed in the analog domain. The development of mixed-signal circuits and architectures that lead to the best deep generative network designs by exploiting unique physics of emerging device technologies is investigated. The project is expected to generate orders of magnitude improvements in energy and delay for deep generative AI models, which will promote their applications and benefit the AI industry.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.