The pleasure of working with a world-class group of visitors this summer
Nearly every summer, the Pattern Recognition group at the School of Computer
Science and Electronic Engineering welcomes new and returning research visitors
from abroad. Over the years, we hosted visitors from Spain, Italy, Netherlands,
Bulgaria, Russia, Lithuania, USA, Poland, Turkey, and even Australia.
Interestingly, our group attracts again and again the same special breed of
people: bright, enthusiastic, incorrigible workaholics.
This summer is no exception. I have the pleasure and the privilege to work
with four wonderful scientists, three from Spain and one from Turkey. The
Spanish team comes from the University
Prof Juan Rodriguez has been coming to Bangor since 2005,
nearly every year. It is great fun to work with Juan!
Razor-sharp, monosyllabic, inhumanly efficient, and
available within 2 minutes after pressing Send on the
e-mail, 24/7, 365 days of the year. Not kidding!
The very start of our collaboration (and lifelong friendship!)
led to one of the most influential publications in my career –
"Rotation Forest". Eh, well, this is neither a forest
in the mountains, nor does it spin around.
Rotation Forest is a
machine learning method where a group of decision tree classifiers
(hence forest) make a prediction about something, and then their
predictions are combined to make a single most accurate
prediction. Like crowd wisdom. And the rotation part is applied
to the data used to train the decision tree classifiers. Think
of the data rotation as mixing up the soil which each tree grows
on. Juan thought of this all; I only played the devil's advocate
and helped putting the finishing touches. Turns out, Rotation
Forest is one of the best state-of-the-art classifiers today.
(I am very proud of this work.) And when I make a remark of how
amazing his work is, Juan only shrugs his shoulders and says,
Over the years, Juan brought to Bangor several members
of his team. This summer we also have Dr Alvar Arnaiz
Gonzalez and Mr Mario Juez Gil. This is Alvar's third
visit to our group. Alvar's qualities bear striking
similarity with Juan's: determined, hard-working,
and brilliant at what he does. And Mario is a young star
in the making. What a pleasure and comfort it is to work
with a team like this! Lucky me!
This summer we chose to work on a popular machine learning
topic whose multiple caveats are often swept under the
carpet: feature selection. Feature selection is about
finding out which are the most important descriptors
that distinguish one group of objects from another.
This work was prompted by a recent collaborative study I did
with the School of Sport, Health and Exercise Sciences at
Bangor University. The task was to create a (preferably
short) list of the most important features (factors)
which discriminate between
athletes who were Great British Olympic and World Champions
("Super-Elite") and international athletes who had not won
major medals ("Elite")
. The features were gathered from the
developmental biographies of 32 athletes, 16 Super-Elite
and 16 Elite. Alright, the data set is too small! Agreed,
but there aren't any more to add there! Can we offer a
reliable list of features from such a small sample?
The same question has been raised more than once about selecting
important genes in cancer research. Using the same (small sample) data,
different research groups came up with completely different lists of
important genes! Beyond the embarrassment, such genuine but
fundamentally flawed analyses may cause damage and even cost lives.
Together with my visitors, we prepared a cautionary tale about
feature selection from ‘wide’ datasets which illustrates the
perils of the current practices. Who knows what the chances are
to get this published? Warnings often get drowned in the stream
of over-enthusiastic (and much of it possibly misguided) novelty
research. But it was a great fun working on the project.
Women can multitask. Fact! Along with Juan, Alvar, and Mario,
this summer I am working on a very different project with
Mr Emre Kavur from The
Dokuz Eylul University
in Izmir, Turkey.
Emre is staying in Bangor for a whole year. Since the beginning
of his visit he has been a super-star. When the Spanish
visitors arrived, Emre instantly found kindred spirits in
them – they all work all hours of the day, and results
appear as if by magic!
Emre's work is on segmentation of organs in the abdomen
from 3D images coming from CT scans or MRI scans.
You would think this problem has been solved many
times over by now. It turns out that the precision of
the segmentation is still not enough for delicate
medical procedures. Emre is a part of a team in
Turkey who organised international competitions
CHAOS) inviting the best players in the world to
cross swards in this challenging problem.
And, of course, these days the good old segmentation methods have
bowed out, and deep learning neural networks have taken over
completely. Emre and I are working on a new segmentation method:
an ensemble of deep learners. Fancy! But this time, unlike
Rotation Forest, we cannot afford to train hundreds of
segmenters (deep learning neural networks) and use them as the
crowd supposed to impart wisdom. We can only use the 'expertise'
of a few such segmenters, which puts the responsibility on
the method for combining the 'expert opinions'. Right up my street!
But it is not only work, as you might have guessed from the photos.
Bangor and its surroundings are lovely this time of the year.
I like to take my visitors to see the gorgeous mountains,
gardens, castles and beaches.
This reminds me of a story a from years ago. It was a glorious,
sunny day at the end of June. I had a visitor (let’s call her
Ivy) whom I took for a walk to Newborough beach. Ivy and I
were enjoying the sunrays, leisurely walking on the wide sandy
beach towards Llanddwyn Island, deep in conversation about
pattern recognition. Before we knew it, the tide
surreptitiously crept in, and we found ourselves stranded
on a shrinking sand island, 50m away from the nearest dry
land. Ivy's eyes widened in disbelief and her jaw dropped
at my invitation to undress herself on a busy beach. But
she quickly got the idea watching me swiftly getting out
of my shoes, socks and trousers, folding them, and placing
them on top of my head, followed by my bag, phone, and camera.
She copied the move, and soon the spectators on the beach
witnessed the procession of two half-naked, serious
lady scientists marching through four feet of water
towards dry sand, balancing precious possessions on their
heads. Ah, it was funny. By the way, Ivy still talks to me.
And the seminar she gave that afternoon was splendid.
You might have thought I had given up trying to drown
my visitors there after this incident? Oh, no!
Newborough beach is so magnificent in any season that
I can’t resist showing it off to my colleagues from abroad.
When will I learn to check the tide times first?
Juan is taller. He only needed to roll up his trousers.
This summer, the five of us went to the woods near
Betws-y-Coed. Of course, we had to wear party hats!
Mario said we looked like unicorns.
On our way back, I swerved the car towards the curb
to let a police car with a flashing light pass by
(the curse of narrow roads). When we started again,
Emre said "Stop the car at once! You have about 45
seconds to do so!" Utterly puzzled, I found a tiny
parking at the back of a shop, just a few meters
away, and stopped the car. We all jumped out and
witnessed the front left tyre quickly deflating
with a fierce "ssssssssss".
What happened next was
like watching F1 racing crew members changing a
tyre in a pitstop! Juan and I were leaning on a
wall while Emre was directing the operation,
and Alvar and Mario were working in amazing
It looked as if they had been changing
car types all their lives. Science, eh?! I couldn't
resist sneaking a few photos of the action.
Summers are my research heaven, especially with a
team like this! I fondly remember all my visitors
and all the thrilling and rewarding work we did
together. I hope they enjoyed Bangor as much as
I enjoy having them year after year.