17 December 2018

Could machine learning save people from drowning?

By Robert Graboyes

Saving a child from drowning and saving a child from cancer are similar in their impact on the children, their loved ones, and society. We classify the second, but not the first, as healthcare — which is a reminder that the borders of healthcare are somewhat porous and arbitrary. I was reminded of this a year or so ago when a Twitter acquaintance, Gennady Simanovsky, emailed to say he had started a company (now called Dolphin Vision) to develop high-tech ways of preventing drowning.

By chance, I had recently read a haunting Slate article, “Drowning Doesn’t Look Like Drowning,” explaining that, contrary to the Hollywood version of a flailing, screaming victim, drowning is actually an eerily quiet, subtle, systematic process. Simanovsky’s company hopes to reduce that process to machine-readable algorithms.

The Slate article explained that the neurological response to the onset of drowning prevents victims from calling for help, waving to onlookers, or reaching for rescue equipment. They remain upright in the water, their mouths repeatedly sinking below the surface and then rising above. Their arms extend laterally and press down on the water. Kicking is minimal. And, most ominously, “Unless rescued by a trained lifeguard, drowning people can only struggle on the surface of the water from 20 to 60 seconds before submersion occurs.”

Dolphin Vision’s idea requires an intelligent machine to understand the concept of water, scan its surface, analyse movement within the water, and identify a swimmer in distress. As is often the case, the company’s developers are an eclectic assortment of experts whose prior experience has little or nothing to do with the subjects at hand — computer vision, the interaction of water and light, or water safety.

On a practical level, Dolphin Vision aims to develop a small, inexpensive electronic device to continually scan a body of water — say, a swimming pool — and its surroundings. It will observe all the people around and in that water. Using machine learning (specifically neural networks), the device will teach itself to identify when one of those people is in danger of drowning, even before the likely victim is aware that a problem exists. Then, it will fire a warning to a safety enforcer (such as a lifeguard or private pool owner).

My knowledge of Dolphin Vision comes from conversations with CEO Simanovsky and CTO Brian Tervil. Their idea is at an early stage of development, and I can’t attest to the validity of the technology or the likelihood it will reach the market. But its goal is compelling and well worth watching.

While I have never met Dolphin Vision’s creators and have no stake in the company, I did make one significant contribution to the enterprise — the company name. Simanovsky has corresponded with me for three years on various topics. In 2017, he thought of naming the company after an ancient sea creature with unusual eye structure and wrote to ask how the prospective name would sound to an English-speaker.

For me, the name was obscure and hard to pronounce. But, I suggested that a different sea animal — the dolphin — is universally recognised, has an easily pronounceable name, and has been associated for millennia with preventing humans from drowning. (Plus, dolphins’ sonar-like capability gives them a rather high-tech air, as well.)

Soon thereafter, Simanovsky told me his company took my suggestion and would, indeed, be known as Dolphin Vision. While it won’t earn me a single dollar, I do enjoy having a modest emotional investment in the company.

Inspiration and Innovation

A recurring pattern in technology is that of innovators from one field introducing breakthrough ideas into other fields. (Jeff Bezos was an investment banker before founding Amazon. The Wright Brothers were bicycle mechanics. Elon Musk co-founded PayPal before branching into automobiles, rocketry, tunnelling, and more.) Like so many innovative companies, the inspiration lies in an emotional or traumatic occurrence.

Simanovsky’s biography shows his origins in Belarus and a career involving printing technology, notably with a subsidiary of Hewlett-Packard. A few years back, he started a company that converted family photographs into three-dimensional statuettes via 3D printing.

The inspiration for Dolphin Vision came during a beach outing. Simanovsky saw a man beginning to drown in the Mediterranean. He helped organise a group of beach-goers into a rescue team and pulled the man from the sea before it was too late. Afterwards, he began wondering how technology might prevent such tragic and near-tragic occurrences.

But Simanovsky’s background was in business, and not in computer vision or water safety. So, he recruited Tervil, a PhD candidate in mathematics at the University of Haifa — who also has no background in computer vision or water safety. But Tervil understood Simanovsky’s goal, decided his mathematical skills were up to the task, and thought Dolphin Vision would provide an enjoyable contrast with his graduate studies.

Tervil is originally from France. His doctoral research at the University of Haifa focuses on symplectic topology — an esoteric branch of mathematics that, suffice it to say, has nothing to do with computer vision or water safety.

Recently, I talked with Tervil, and his observations form the bulk of what I know of the Dolphin Vision enterprise. We spoke at length about the idea behind the company, how its device might work, and what challenges are inherent in building such a system.

The first problem, he said, is strikingly basic. Computers don’t really understand the concept of water. It’s remarkably difficult, he said, to explain to a computer what water is, how it behaves, and how objects immersed in water appear to the eye or to a camera looking down on the water from above. How water moves and the way light reflects on and refracts through water confound the computational processes at the heart of computer vision and recognition. And, Tervil added, the scientific literature on these topics is far smaller and less advanced than one might think.

Recognising Children, Recognising Drowning

Tervil describes Dolphin Vision’s analytical challenge as a two-layered problem. The first layer is determining who around the swimming pool is a child. The second is recognising who in the pool is in danger of drowning.

The first layer alerts the safety enforcer when a child approaches swimming pool unaccompanied. At that point, the enforcer can determine whether the child ought to be there and, if not, whether to take action. By means of machine learning the device will teach itself to recognise children.

The second analytical layer is learning to identify drowning so early that the victim may not yet realise that he is in trouble. (This video suggests how a computer might “see” a swimmer.) The device will signal the safety enforcer. This could be a lifeguard whose attention is divided among many swimmers under his watch. Or it could be a homeowner inside his house.

Recall the key datum from the Slate article: “drowning people can only struggle on the surface of the water from 20 to 60 seconds before submersion occurs.” To function, Dolphin Vision’s device must be able to recognise these signs rapidly and instantly alert a safety enforcer. Tervil stressed, “The goal is not to replace the lifeguard, but to give him help.”

Going Forward

As with many technology startups, Dolphin Vision faces three hurdles: science, engineering, and business. They must crack the optical science and its associated mathematics. They must house that science within a device that is small, affordable, and reliable. And then there’s the task of bringing the finished product to market.

Simanovsky and Tervil reviewed the scientific and technical literature on the intersection of computer vision and the optical characteristics of water. Based on early findings, Tervil constructed a small prototype for demonstration purposes. This enabled them to secure seed funding from the Israel Innovation Authority (formerly Office of the Chief Scientist) and from some private investors.

The company opened for business in March 2018, at which time they hired a chief scientist, Michael Rudek, who, among other things, would write Dolphin Vision’s software.

Tervil says one of the company’s strengths is the fact that the three principals are from different countries, with three different science and engineering cultures. Simanovsky immigrated to Israel from Belarus, Tervil came from France. Rudek is originally from Poland.

Tervil says that to get past that scientific hurdle, the company will have to hire two additional staff members—a mathematical researcher and a specialist in computer vision. From the moment the additional staffers are in place, he believes, it will take five to six months to develop a physical proof-of-concept. Then will come the engineering and commercialisation processes.

Assuming Dolphin Vision reaches those stages, the technology is likely to have alternative uses. As is often the case, new technologies have a way of finding purposes beyond their original aims. Even at this early stage, two other applications are under discussion.

First is in the area of maritime and homeland security—monitoring the movement of objects in navigable waters. (Just recently, Dolphin Vision competed in the US Coast Guard’s Ready for Rescue Challenge. The company has also entered another innovation competition for 2019—MassChallenge.)

The second application is the technology’s ability to observe and analyse the movements of fish. (This video gives an idea of how the device might “see” fish.)

While knowing little about Dolphin Vision’s chances of commercial viability, here’s wishing them the best of success.

Robert Graboyes is a senior research fellow with the Mercatus Center at George Mason University, where he focuses much of his work on healthcare innovation.