Machine
learning is a branch of artificial intelligence where computers can learn from
data, recognize patterns, and make decisions without being explicitly
programmed. It is already used in many ways today, particularly online, for
example in predicting search terms or tailoring adverts to user tastes. As well
as in industry and academia, science is an area in which machine learning
techniques are increasingly being adopted in order to process and analyse a vast
amount of data.
A new
frontier
This subject
was formally addressed for the first time within ICES by the recent Workshop on Uses of Machine Learning in Marine Science (WKMLEARN). Its drive was to explore
potential niches for machine learning within fisheries and marine science as
well as ICES work, where the technology could be of the most benefit, and where
it is already used in the marine sciences. WKMLEARN looked at both machine
learning as well as a subset known as 'deep learning'. Inspired by the
structure and function of a human brain, deep learning is a 'deeper' form of
learning that uses multi-layered artificial neural networks to solve complex
nonlinear problems like recognizing objects in images. It has performed closer
to the level of humans than traditional machine learning.
New opportunities
For marine
science the advantages of machine learning range from modelling and prediction
of climate data and obtaining fishing patterns from satellite images or AIS data, to taxonomic recognition of biological samples, and analysing text,
images and video. Harnessing this technology can be especially fruitful in an
age of exponential growth of data, used to feed the algorithms behind these
systems.
Many
opportunities and challenges are linked to the fact that, as a catalyst for
machine learning, much data and computing is being moved into the cloud. Although
free from hardware restraints and with data secure for reuse, there are issues
such as balancing matters of intellectual and proprietary rights while allowing
consumer access. As ICES goes down a path of reusability, clear data licencing
becomes an issue, and despite advancements, there is still much to do.
Many
research projects will harness the power of machine learning in coming years,
and this will provide opportunity to improve the skills and tools available to
ICES community and the potential speed of science, data, and assessment
projects.
Benefitting advice
Ideas were offered on where new technologies might replace traditional activities across both science and advisory spectrums. Several immediate opportunities in the single species advisory process were identified, in analysing samples and preparing data. These actions can be performed and reproduced quickly. Datasets on fish age, nephrops burrow counts, and acoustic survey interpretation represent are good examples, as there is a frequent availability of annotated images. Despite the benefits, there would be a continued need for human insight and knowledge, which can’t be automatized, to interpret assessment outputs.
Image
recognition – such as predicting fish ages from otolith images – is a task
which can be done consistently while dealing with a large volume of data. Such
learning would not be affected by external factors. These benefits are typical
of 'supervised' systems, where both model input and output are
known, and the machine needs to be taught how to get from one to the other.
Humans can quality control the process and set the parameters.
Longer-term
possibilities include adding ecosystem and environmental information to the
advisory process and products like ecosystem assessments and overviews. This
requires a deeper understanding of how machines could perform these tasks, as
well as training. Examples here are automated identification of plankton for
estimating biodiversity, benthic habitat classification, and environmentally
influenced forecast predictions.
Human impressions
Shaheen
Syed, WKMLEARN co-chair, reflected on the workshop.
“It is nice
to see some many interesting projects using the power of machine learning to
assist in some of the current fisheries and marine science challenges. Although
we see great improvements, there is still a big human component and need for
massive amounts of labelled training data to help these models 'learn' to make good predictions. A fully stand-alone
machine learning system is still far away but we are on the right track."
For Ketil
Malde, who chaired the workshop along with Syed, one of the initiative's main
strengths was in pooling human resources.
“I'm really
impressed that we managed to collect people who apply various machine learning
technology in many areas over almost the whole fisheries process," said
Malde.
“But these
projects are small and isolated from each other. We don't have good common
grounds to meet up, learn from each other, and maintain knowledge. We need
to attain critical mass for expertise and deployment and also prod the people
who are sitting on the data to get it in shape so we can start
using it."