Files and links (1)
CC BY V4.0, Open Access
Thesis
Active and Passive Learning for Training Data Generation in Transistor Modeling
Washington State University
Master of Science (MS), Washington State University
01/2021
DOI:
https://doi.org/10.7273/000005396
Handle:
https://hdl.handle.net/2376/125179
Abstract
Modeling and predicting transistors’ behavior with machine learning is a high-dimensional regression problem. It has a scenario that unlabeled data is meaningful, the query has equal cost and features are numeric. During the modeling, a situation that model performs poorly in regions with high variation in output space was observed. To gain a better performance with smaller data size, a query synthesis active learning strategy and a passive learning strategy are proposed to allocate more samples in high-variant regions. The active learning approach is oriented by model performance. In this strategy, the scope would be split into multiple disjoint cells in input space to estimate and optimize their performance individually. By this way, the samples would be mainly taken in poorly performing cells. The passive learning, called Matrix Smooth, is designed based on a different idea: Since the high-variant regions can be located according to model performance, the poorly performed regions for model can also be identified by finding high-variant regions. Thus, this strategy is oriented by data analysis. It supposes the variation of output space can be reflected by the geometric angles between samples in both of input and output space. By stipulating the initial samples are sampled evenly in input space, the features are orthogonal to each other, then the angles would be only influenced by output space. New samples would be iteratively taken near the sharpest angle to smooth the matrix. Compared to active learning, passive learning avoids training model in each sampling iteration, thus significantly reduces the computational cost. The model also takes advantage of determined data size for hyper-parameter tuning. The experiments on Support Vector Regression and Deep Neural Networks show that both proposed strategies outperform evenly sampling. Especially, the Matrix Smooth is found having the ability of avoiding overfitting on Deep Neural Networks.
Metrics
Details
- Title
- Active and Passive Learning for Training Data Generation in Transistor Modeling
- Creators
- Da Liu
- Contributors
- Xinghui Zhao (Advisor)Scott Wallace (Committee Member)Xuechen Zhang (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- School of Electrical Engineering and Computer Science
- Theses and Dissertations
- Master of Science (MS), Washington State University
- Publisher
- Washington State University
- Number of pages
- 82
- Identifiers
- 99900592256701842
- Language
- English
- Resource Type
- Thesis