Exploring Deep Machine Learning for Automatic Right Whale Recognition and Novel Drug Design
1. http://multimedialab.elis.ugent.be
Ghent University – iMinds, ELIS Department/Multimedia Lab
Ghent University Global Campus – Center for Biotech Data Science
Mijung Kim, Wesley De Neve, Peter Lambert
FEA Research Symposium 2015
{mijung.kim, wesley.deneve, peter.lambert}@ugent.be
EXPLORING DEEP LEARNING FOR AUTOMATIC
RIGHT WHALE RECOGNITION AND NOVEL DRUG DESIGN
9th December | Ghent | Belgium
Automatic Right Whale Recognition
Identifying right whales! Hmm…
How about using deep learning?
If you have many photos, then I can train
a deep convolutional neural network for
identifying the individual right whales.
However, the low number
of photos and the many
classes make this
challenging!
Still, let us give it a try…
Novel Drug Design
Hello
Mijung!
How is your
research going?
Good! I am
currently exploring
DEEP LEARNING, a
novel technique in
the field of machine
learning.
I recently had a talk with Jasper, a marine biologist. He is doing research on
North Atlantic Right Whales, an endangered species with fewer than 500
animals left. So, to monitor the health and the status of the remaining
population, Jasper and his colleagues take photographs of right whales
during aerial surveys, and then manually identify the animals
photographed. However, manually identifying right whales is a time-
consuming process, requiring special training. Mijung, would it somehow
be possible for you to automate the identification process?
Look at this! By giving
photos as an input to a
convolutional neural
network, it will
automatically extract
those visual features
that make it possible
to identify each
individual whale.
Yesterday, I met Jozef. He tried to develop a
new rejuvenation drug, but his chemical
experiment ended up in failure because he
could not predict the side effect of the
chemical synthesis. If he would have
known about my research, he would have
been able to predict the explosion at least.
In particular, I am currently investigating the use of multi-task neural networks to design novel drugs.
These networks do not only consider the main task but also related tasks. So, by making use of these in-
silico networks during drug design, I can predict side effects like the toxicity of combinations of chemical
molecules, without having to make use of in-vivo or in-vitro experiments.
Task 1
Task 2
Hidden
Layer
Hello
Wesley!
Interesting!
Mijung, can you
apply deep
learning to other
biotech problems
as well?
Yes! I can give
you a real-world
example.
Whale 1
Whale 2
Whale 3
Whale 4
convolution +
nonlinearity
max pooling
convolution + pooling layers
vector
Fully connected layersfully connected layers multi-class classification
Whale 1
Whale 2
Whale 3
Whale 4
multi-task
neural network