Cognitive Architecture for Circuit Design Innovation with
Topological Feature Extraction and Causal Information Mining
March 9, 2018
Light Engineering room 250
Advisor: Prof. Alex Doboli
Innovative ideas and applications are emerging and published in electronic documents. It is arguably important to integrate the newly proposed techniques and implementations for EDA tools for analog circuit design. However, it is rather cumbersome to constantly review the cutting-edge techniques and solutions in the literature. Moreover, the design knowledge in high performance analog circuit design is constantly innovating, especially in widely-used applications such as health care, mobile telecommunications. In order to efficiently incorporate the knowledge in electronic form, it is necessary to have comprehensive representation of the meta-knowledge. Furthermore, in-depth analysis on the causality of the extracted meta-knowledge is needed in order to utilize it.
In this thesis, a cognitive architecture is presented for creative problem solving in analog circuit design. The architectural modules attempt to replicate cognitive human activities, like concept formation, comparison, and concept combination. The architecture uses multiple knowledge representations organized using topological similarity and causality information. More specifically, it learns analog circuit designs and analyzes them for design knowledge identification and reuse. Two main techniques, topological feature extraction and causal information mining, are presented for the cognitive architecture. The former extracts the most straightforward features of a circuit-topological structures. This simulates the process of a human learning the schematics of circuits. The latter mines the causal relations within the circuits including performance tradeoffs, design preferences and constraints. Comprehensive causal information mining techniques make the cognitive architecture behave close to an expert designer.
While topological feature extraction and causal information mining are presented as two major modules of the architecture, they are also linked for mutual benefits. Causal information helps justify topological feature extraction for unsupervised learning, which in turns provides guidance for mining causal information. The combination of two techniques makes a complete cycle that runs without human interference in ideal situation. As a prototypical experiment, A large amount of experimental data is presented in this thesis. A total of 34 state-of-art OpAmp/OTA circuits are considered. We also show how the system can be scaled to even larger input sizes.