Hongjin Yu
Software Development
«
Convolutional Kernel Optimization for Deep Neural Networks using
Constructivist Augmented Machine Learning (CAML) Methodology
Deep neural networks have beaten I was always interested in how neural
Go and more recently Starcraft II. Despite approach the subject because it seemed
humans at image recognition in the game
these achievements, the inner workings
of these networks have remained a black
box. The first half of my thesis focused on
understanding the hidden layers through
visualization. I gave the network different
stimuli and observed how different regions
in the network reacted. In the second half
of my thesis, I used the patterns found in
the visualizations, combined with human
knowledge and intuition, to manually
adjust the networks to improve them.
My results showed that humans could
predict where networks converged. With
further research, this could greatly reduce
the amount of time and data needed for
training networks.
100
SOFTWARE DEVELOPMENT
networks worked but never dared
daunting. This project gave me the
perfect opportunity to challenge and force
myself to study something that I’ve really
wanted to learn. When I actually dug into
the project, I found out that it really was
not that complicated. Fear was the only
thing that made it seem impossible. After
completing this project, I now have the
confidence to take on more challenging
problems in the future.