 |
|
|
| (Seminar) Bayesian tensor network: towards efficient and interpretable probabilistic machine learning |
|
CAS Key Laboratory of Theoretical Physics
|
|
Institute of Theoretical Physics
|
|
Chinese Academy of Sciences
|
|
Seminar
|
|
Title
题目
|
Bayesian tensor network: towards efficient and interpretable probabilistic machine learning |
|
Speaker
报告人
|
|
|
Affiliation
所在单位
|
首都师范大学 |
|
Date
日期
|
2020年11月2日10:00-11:00 |
|
Venue
地点
|
ITP South Building 6420
|
|
Contact Person
所内联系人
|
张潘 |
|
Abstract
摘要
|
Developing novel machine learning models with both high interpretability and efficiency is an important but extremely challenging issue. In this work, Bayesian tensor network (BTN) is proposed by combining Bayesian statistics with tensor network (TN), which captures the conditional probabilities of exponentially many events efficiently with polynomial complexity and meanwhile retraining high interpretability as a probabilistic model. BTN is tested on classifying images of hand-written digits and fashion articles, where the classification tasks are mapped to the problems of capturing the conditional probabilities in an exponentially large sample space. Impressive performance using simple loop-free structures are demonstrated with insignificant over-fitting. Furthermore, BTN can be used to as a module to construct novel end-to-end models by hybridizing with neural networks. |
|
|
 |
|
