A Survey on Conditional Convolution Neural Network

PhD Qualifying Examination


Title: "A Survey on Conditional Convolution Neural Network"

by

Mr. Ningning MA


Abstract:

Convolution neural network is a robust and fundamental method in computer
vision tasks. Researches have been conducted to improve the two main
aspects of CNNs: the effectiveness and the efficiency. Deep CNNs are
proposed to improve the effectiveness, light-weight CNNs are proposed to
improve the efficiency. Both deep CNNs and light-weight CNNs are standard
CNNs which share the same convolution kernels and the same network
structure for all the image samples across the dataset. Nowadays, as a new
technique, conditional CNNs establish the CNNs (kernels or network
structures) conditioned on image samples. In this survey, we focus on the
convolution layers and the activation function layers since they are the
two most common layers in CNNs. First, we survey the standard CNNs and
conditional CNNs. Second, we review the existing attention mechanisms in a
generalized conditional convolution formulation. At last, we survey the
scalar activation functions and the conditional activation functions. The
research has also compared standard CNNs and conditional CNNs in many
aspects, shown that conditional CNN is a promising future direction.


Date:                   Friday, 14 February 2020

Time:                   1:30pm - 3:30pm

Zoom Meeting:           https://hkust.zoom.us/j/910034874

Committee Members:      Prof. Long Quan (Supervisor)
                        Dr. Qifeng Chen (Chairperson)
                        Dr. Xiaojuan Ma
                        Prof. Chiew-Lan Tai


**** ALL are Welcome ****