Page 49 - Undergraduate Research Opportunities Program Proceedings 2017-18

School of Science
Department of Physics

Neural Dynamics

Supervisor: WONG Michael Kwok Yee / PHYS
Student: TAI Wai Ting / PHYS-IRE Course: UROP2100, Fall

Neural network is most commonly used for the classification of discrete data, but its behavior for continuous data has
also been a subject of wide interest, as it can shed light on the mental representation of continuous data. This project
aims to study the ability of a neural network to handle images in different orientations. A neural network is trained with
images of rotating chairs and is able to learn a continuous distribution of the image based on its orientation. Afterwards,
networks are trained with images of rotating sofas in different colors, and the results suggested that the network is able
to generalize to some extent to different colors, showing that the distribution is also applicable to images of different
colors.

Supervisor: WONG Michael Kwok Yee / PHYS
Student: YUAN Weijun / PHYS Course: UROP2100, Fall

Understanding how the brain works is one of the most challenging issues in modern scientific research, but many
questions remain unanswered. One possible approach to address this problem is to utilize the tools of deep learning to
simulate the functions of the brain. The purpose of this study is to realize the mechanism of rotational motion
recognition in human brain using the artificial neural network. It is shown from deep visualization that the deep neural
network can extract the features of an image with hierarchical structure, which is very similar to the visual system. The
VGGNet trained with Imagenet data is used to mimic the visual system and a feed-forward neural network is designed to
simulate the function of rotational-motion recognition.

39