to associate a visual stimulus with a food reward. The stimulus
would be presented to the ish on a computer screen, or as a
laminated paper print-out, and the ish would swim up to the
stimulus and bite it. As soon as it had done this it would receive
a food reward.
When the ish had made this association and performed the
task quickly and reliably, a second distracting stimulus was
introduced that differed from the irst. The ish’s task was then
to identify the original stimulus in order to get the food reward.
More distracting stimuli could be added to increase the power
of the experiments, and the stimuli could be made more or less
similar to test the sensitivity of the ish to slight changes.
The amazing thing is that the ish learned to associate a food
reward with a visual target within just days of capture, and they
learned the task so well that it became possible to deliver the food
with a delay and in a different part of the aquarium to where
the stimuli were displayed. This anticipatory behaviour indicates
higher learning capabilities, as the stimulus is not treated as a
substitute to food but as part of a procedure that in the end
will lead to the supply of food.
So what did we ind out? We trained several ambon
damselish to associate a food reward with a face stimulus
belonging to another ambon damselish individual, and trained
several other ambon damselish to associate a food reward with
a face belonging to a lemon damselish individual. Both were
then tested against novel ambon damselish faces. We were
excited to ind that all ish were able to reliably discriminate
species as well as individuals from the same species on the basis
of their facial patterns alone. So ish, like us, use facial features
to discriminate individuals!
Our next question was which features were important for face
discrimination. Could the ish only perform the task when the
patterns were displayed in ultraviolet colours?
To test this, we repeated this experiment using black and
white patterns instead of the ultraviolet patterns, and found
that the important features were the shape of the patterns rather
than their colour. From this point onwards we have been
working with black and white or greyscale face images, which
makes our lives a lot easier as we can now display the faces on
a computer screen (which cannot produce ultraviolet colours)
rather than having to print them out and go through the labour-
intensive process required to produce ultraviolet stimuli.
This series of experiments, together with technologies derived
from machine learning and computer vision, has set the scene
for a large number of questions we have been addressing over
the past few years. We developed morphing software that allows
us to create smooth continua of intermediate stimuli between
any two original faces. This allows us to test the limits of the
discrimination ability of the ish, as we can make the distracting
stimuli increasingly similar to the original stimulus.
At present we are investigating which features within the
facial patterns the ish are using. Humans pay most attention
to the eyes, nose and mouth when performing human face
discrimination. We are using a combination of machine learning,
computer vision and behavioural experiments to test if this is
similar for ish. Our initial results show that the eyes are partic-
ularly important for ish as well.
Another line of our work is to test whether ambon damselish
and archerish can discriminate human faces as well as their
own faces. Our initial results show that this is indeed the case.
Overall, our work to date suggests that ish have the ability
to learn visual discrimination tasks within just a few days from
capture, and to a high level of accuracy (>80%). They learn to
discriminate ish faces and human faces as fast and as well as
they learn to discriminate objects such as squares and triangles,
or differently coloured shapes. They show anticipatory behav-
iour, and have the capacity for lexible decision-making
depending on the speciic conditions of a particular experi-
ment.
Our results demonstrate that ish have much better cogni-
tive abilities than is generally assumed, which makes them a
great model for the study of visual learning, processing and
perception in general and speciically for the evolutionary origins
of visual abilities that are thought to require a cortex, such as
face recognition.
So next time you pass a ish in a tank, just recall that they may
well be watching you.
Ulrike Siebeck is a senior research fellow at the University of Queensland’s School of
Biomedical Sciences, where she heads up the Visual Neuroethology Laboratory. Guy Wallis
is an academic in the Centre for Sensorimotor Behaviour in the School of Human Movement
and Nutrition Sciences at the University of Queensland.JUNE 2016|| 17Ambon damselfish performing visual discrimination tasks. The
main image shows a colour discrimination task: the fish has to
swim up to and touch the colour card to which it was previously
trained. The inset image shows a facial pattern discrimination task:
the facial patterns of an ambon damselfish and a lemon damselfish
are shown in white (corresponding to the UV facial patterns of a
fish) on a black background. In this test the fish is correctly
identifying the facial pattern of another ambon damselfish.