Neuromorphic computing and other nature-inspired methods for hardware and software design

本研究涉及使用人工智能（AI）技术在射频（RF）下操作的模拟电路的自动尺寸。人工智能最近的成功在很大程度上是由于存在大量用于训练的历史数据，以及主要需要考虑的一对一关系。这些特征在工程设计中是不典型的，因为工程设计中的问题往往是不确定的，训练的例子也很少。还有一个要求是，解决方案必须能够解释透明、信任和帮助决策，并且在电子设计自动化（EDA）的情况下，它们必须考虑到隐含的设计约束。提出的研究将为解决这些问题的模拟射频电路的自动尺寸制定一个框架。将进行三项主要活动：创建一个有效的人工神经网络（ANN）架构，用于射频电路的自动合成和元件尺寸确定，该架构准确，可以从少量示例中学习，并处理诸如元件耦合之类的隐式约束。2.找到一种有效的优化算法来补偿RF电路尺寸上的布局，阻抗不匹配和可变性问题等实施影响，使用创建的人工神经网络的输出作为初始启动。3.开发一种方法，根据白盒操作的领域知识来解释所创建的人工神经网络体系结构的操作，更容易进行超参数调整，并允许专家在调整过程中输入。前面的任务将通过神经网络、进化优化算法和模糊逻辑的组合来完成。为了降低人工神经网络的复杂度以满足较低的训练要求，将研究稀疏计算技术，如变大小分区和部分计算，对于现实世界的效果负责，将采用进化优化算法将人工神经网络的输出作为初始解，并使用代理模式技术实现快速收敛，对于可解释性，将使用模糊提取输入输出关系和领域知识增强。所创建的设计方法为模拟电路设计和这些电路的更快设计的声明性方法打开了大门。鉴于无线设备的需求不断增加，其模拟RF前端的复杂性不断增加，EDA工具变得必不可少。此外，它可以很容易地适用于其他非线性设计问题，而不是在本研究中所讨论的问题。最后，参与本研究的学生将学习多学科技能，包括工程学、人工智能和认知信息学。因此，该研究项目也将有助于保持加拿大在人工智能整体领域的领先地位。

This research is concerned with the automatic sizing of analog circuits that operated at radio frequencies (RF) by using artificial intelligence (AI) techniques. The recent success of artificial intelligence is in big part due to the existence of large amounts of historical data for training, and of mainly one-to-one relationships to consider. Those features are atypical in engineering design, where the problems are often undetermined and the training examples few. There is also the requirement that the solutions be explainable for transparency, trust and help in decision-making, and in the case of electronic design automation (EDA), that they account for implicit design constraints. The proposed research will elaborate a framework for the automatic sizing of analog RF circuits that address those issues. Three main activities will be undertaken: 1.Create an efficient artificial neural network (ANN) architecture for the automatic synthesis and component sizing of RF circuits that is accurate and can learn from few examples and handle implicit constraints such as component coupling. 2.Find an efficient optimization algorithm to compensate for implementation effects such as layout, impedance mismatches and variability issues on the sizing of RF circuits, using the output of the created ANN as initial start. 3.Develop a methodology to interpret the operation of the created ANN architecture in terms of domain knowledge for white-box operation, easier hyperparameter tuning, and to allow for expert input during the sizing process The previous tasks will be accomplished through a combination of neural network, evolutionary optimization algorithm and fuzzy logic. To reduce ANN complexity for lower training requirements, sparse computing techniques such as variable-size partitioning and partial computations will be investigated, for real-world effects accountability, an evolutionary optimization algorithm will take the ANN output as initial solution and use surrogate mode techniques to achieve fast convergence, and for explainability, fuzzy extraction of input-output relationships with domain knowledge enhancement will be used. The created design methodology opens the door to a declarative approach to analog circuit design and the faster design of those circuits. Given the increasing needs of the wireless devices keeps increasing, and the ever-increasing complexity of their analog RF front ends, EDA tools are becoming mandatory. Moreover, it can be easily adapted to other nonlinear design problems than the one addressed in the proposed research. Finally, the students who take part in this research will learn multidisciplinary skills, with input from engineering, artificial intelligence and cognitive informatics. Thus, this research program will also contribute to maintain Canada's leader position in the overall field of artificial intelligence.