Recently, I've become aware of a tremendous increase in the use of deep learning and artificial neural networks (ANNs). (See CEVA Accelerates Deep Neural Networks and Biometric fusion-based security meets deep learning for smartphones & tablets.)
I was familiar with the use of ANNs in biometric applications like identifying people in crowds, but I'd not really thought about using this technology in the context of saving endangered species. All this changed when I talked to some folks from MathWorks and NOAA (the National Oceanic and Atmospheric Administration) about their use of ANNs to save right whales.
Right whales are so-named because they were considered to be “the right whales to hunt.” In the early centuries of shore-based whaling (prior to the early 1700s), right whales were virtually the only large whales that were catchable because they often swam close to shore, they are relatively slow swimmers, and — once killed by harpoons — they were likely to float, thereby facilitating retrieval.
As an aside, these creatures are huge; they can grow up to more than 18 m (59 ft) long — the largest recorded thus far was 19.8 m (65 ft) — and they can weigh in at up to 100 short tons.
By 1937, it is estimated that more than 100,000 right whales had been killed and the species was close to extinction, so the world agreed to ban right whaling. Of course, some rogue countries continued hunting — the usual suspects — such as the Soviet Union, which illegally took at least 3,000+ southern right whales during the 1950s and 1960s (it reported taking only four).
No one really knows how close we came to losing right whales forever. Fortunately, under the custodianship of organizations like NOAA, things are starting to look a little less bleak, and there are now around 500 living North Atlantic right whales.
(Source: Dr. Richard Pace, National Marine Fisheries Service)
Part of the conservation effort involves tracking the whales as they migrate their way across the oceans and observing which whales mate together, how often females give birth, who is related to who, and so forth. The least intrusive method for doing this is by taking aerial photographs and then analyzing these images to determine which whales are in the scene.
Until now, the task of identifying the whales has been performed by a small, dedicated group of people. Some of these heroes can recognize all ~500 living whales, but the task is becoming increasingly difficult as the number of right whales grows. At some stage in the not-so-distant future, it will become almost impossible for human observers to be able to correctly identify each whale.
Have you heard of Kaggle? This is the world's largest community of data scientists who compete with each other to solve complex data science problems. The Kaggle community comprises tens of thousands of PhDs hailing from over 100 countries and 200 universities; from quantitative fields such as computer science, statistics, econometrics, maths, and physics; and from industries such as insurance, finance, science, and technology.
The point is that there is currently a Right Whale Recognition Competition running on Kaggle. The folks from MathWorks are sponsoring the competition prize pool, and they are also providing complimentary software to teams participating in the competition (they have an extensive set of ANN tools, libraries, and algorithms along with a tool box for deep learning applications). Meanwhile, the right whale research team at the New England Aquarium has been maintaining a massive photo-identification catalog, and Christin Khan and Leah Crowe from NOAA have undertaken the laborious task of hand-labeling the images to create a one-of-a-kind dataset for use in the competition.
(Source: MathWorks and NOAA/NEFSC/Christin Khan)
This has to be an incredible complex task. First of all you have to isolate the whale from the noisy background (water), plus you have to deal with the fact that there may be multiple animals in the picture — such as a mother and her calf — and these may be close together or even touching.
On the bright side, creating an ANN that can accurately identify right whales has to be a tremendously fulfilling task. I only wish I had the knowledge and skill to become involved. Happily, there is still time for you to form a team and enter this competition, which will end at 11:59 pm UTC on Thursday 7 January 2016. I can't wait to hear more about the winning solutions. Until then, can you think of any other wildlife conservation projects that could benefit from the application of ANNs?