讲座标题:Explanatory graphs for CNNs &
interpretable CNNs
时间:2017年9月29日,星期五,上午10:00-11:00
地点:电信3-412
主持:王新兵教授
Abstract:
Although convolutional neural networks
(CNNs) have achieved superior performance in different visual tasks, the
knowledge representation inside a CNN is still considered as a black box. In
this talk, I mainly introduce two of my studies to enhance the interpretability
of the knowledge encoded in conv-layers of a CNN, i.e., 1) learning a graphical
model, namely an explanatory graph, which reveals the knowledge hierarchy
hidden inside a pre-trained CNN, and 2) end-to-end learning an interpretable
CNN, whose filters in high conv-layers encode semantically meaningful patterns.
1. For the explanatory graph: Considering
that each filter in a conv-layer of a pre-trained CNN usually represents a
mixture of object parts, I propose a simple yet efficient method to
automatically disentangle different part patterns from each filter, and
construct an explanatory graph. Each graph node represents a part pattern, and
graph edges encode co-activation relationships and spatial relationships
between patterns. I learn the explanatory graph for a pre-trained CNN in an
unsupervised manner, i.e. without a need for annotating object parts.
2. For interpretable CNNs: I design an
interpretable CNN, in which each filter in a high conv-layer represents an
object part or a discriminative texture. The interpretable CNN can be designed
with different loses for different tasks. Without given additional annotations
of object parts or textures for supervision, the interpretable CNN
automatically assigns each filter with clear part/texture semantics during the
learning process.
Speaker bio:
Quanshi Zhang received the BS degree in
machine intelligence from the Peking University, China, in 2009 and M.S. and
Ph.D. degrees in the center for spatial information science at the University
of Tokyo, Japan, in 2011 and 2014, respectively. Now, he is a postdoctoral
researcher at the University of California, Los Angeles, under the supervision
of Prof. Song-Chun Zhu. His research interests range across computer vision and
machine learning. Now, he is leading a group for explainable AI. The related
topics include explainable neural networks, explanation of pre-trained neural
networks, and unsupervised/weakly-supervised learning.
Homepage: https://sites.google.com/site/quanshizhang/