How the Sumatran Rhino Taught Me to Rethink AI
Having worked in the most beautiful and unforgiving environments on every continent, I’m always a little surprised how anxious I still get when deploying one of my new technologies. The first real field test of a prototype is always hard for an inventor . . . Will it work like it did in the lab? What did I miss? I sure wish all these people weren’t watching as I flip the power switch . . .
This time the project was even more intense. I was deploying a custom 3D-scanning rig that I built from $60,000 in parts to try to create a digital copy of one of the world’s last remaining Sumatran Rhinos at a sanctuary deep in the humid Indonesian jungle. This could be our one shot to capture a digital copy of the rhino that would help tell its story, or in the worst case provide a full digital record of the animal if conservation efforts failed. To up the stakes even more, there was a National Geographic crew filming every step of our efforts.
My career has been a long series (a few decades now, but I try not to think about that too much!) of these types of sprints. Months of lab development of high-tech systems for conservation, exploration, or anti-poaching security applications punctuated by intense field deployments. Having a background in Computer Science and Artificial Intelligence from the Massachusetts Institute of Technology (MIT), I did not initially see myself ending up in a career where I was building technologies and algorithms for exploration and conservation. Working with National Geographic gave me a chance to merge my passions for computer science and AI with conservation, exploration, and storytelling.
Being an Explorer and Fellow at National Geographic is synonymous with incredible adventures. Unlike most jobs, it really is as cool as it sounds. I have been on the deck of the Titanic 12,500’ deep in a 3-person submarine while battery technology I helped develop powered robots exploring the passenger cabins inside the wreck. Last year I helicoptered over Mount Everest as we captured a high-resolution aerial LIDAR map of the entirety of the world’s highest glacier. With our conservation partners I flew cameras strapped to the wing of a twin-seat aircraft over the Democratic Republic of the Congo on the front line of the fight against elephant poaching. I have spent many hundreds of hours in the pitch black underworld on missions like scuba diving in flooded caves to scan Mayan human-sacrifice victims and jamming my body through a vertical crack that is only 7-inches wide in some places, where falling would leave you wedged and broken, in a South African cave system to search for pre-human burials with ground penetrating radar.
Taken as a whole, this has left me inspired and excited for the role that technology and innovation can play in sharing and protecting the world’s most beautiful treasures while terrifying me with how fragile the things we care about really are.
For my work I was recently named the 2020 Rolex National Geographic Explorer of the Year. Each year this prestigious award is granted to an individual whose actions, achievements and spirit personify leadership in exploration and storytelling. The previous winners include filmmaker James Cameron, underwater photographer Brian Skerry, Steve Boyes and his Okavango Wilderness Project team, and environmental anthropologist Kenny Broad. These days the National Geographic Society mission - to use the power of science, exploration, education and storytelling to illuminate and protect the wonder of our world - is more important than ever. I am incredibly honored to be a part of that mission.
While I was digitizing the Sumatran Rhino, I rendered a preview visualization on my laptop back at the lodge where we were staying in Indonesia. As I moved the 3D representation of the animal around on my screen, I was struck by how incredibly realistic it was. When I stopped moving my mouse, I could hardly tell that the view was not a photo. Having an artificial intelligence background, as well as a philosophical bent, I started thinking what is a “real” image? I certainly think of a photo I take with my phone as a real image, but what is that really except a bunch of numbers representing pixel values? If my synthetic digitized rhino is indistinguishable from a photo, is it as real as a photo? I realized that if I can create 3D models that look real to me, I can use these images to train AI systems. You see, data is the Achille’s heel of artificial intelligence. A good AI model needs a LOT of data, but in some of the most impactful applications - such as conservation, anti-poaching, security, and medical imaging - good data is hard to come by.
This idea led to my current adventure in founding Synthetaic, an AI and synthetic data company that is focused on some of the most high stakes use cases for artificial intelligence where limited sample islands have prevented high-quality predictive modeling. In many areas, there simply hasn’t been sufficient data to train networks effectively, for both still and moving images in particular.
Photos of illegal poachers are scarce. So too are those of late-model Toyota pickups sporting extremist insignia. Real-time feedback for neurosurgeons in the operating room has been held back by limited image repositories of rare brain cancers. Racial bias runs rampant in facial recognition because of prevalence issues in training data.
I started Synthetaic to answer what I see as the two biggest questions in AI. What if edge cases no longer existed? What if training data was no longer a constraint?
Since we started Synthetaic, we have expanded the concept as we realized that by combining both 3D modeling and novel generative AI we could grow data faster, cheaper, and in a way more aligned with training state-of-the-art AI than any other data synthesis technique. These days we are generating human cancer microscopy images for inter-operative brain surgery decision making, growing chest x-ray images for COVID-19 detection, synthesizing aerial imagery for conservation projects, and creating data for several unsolved security intelligence needs. In each of these cases the data is indistinguishable from real images, but nearly instantly generated on our servers, and we can use this data to train AI models that outperform the current state-of-the-art.
We need technology as a force multiplier for the world’s hardest problems. Even what we now see as simple technologies, like the pulley or electric power have radically increased our capabilities as individuals and as a species. I believe that artificial intelligence, when taken to its full potential, will be such a tool. It will help out numbered rangers to protect vast areas from poaching. It will help doctors diagnose medical problems earlier when more treatment options are on the table. It will help keep us safe in an uncertain world. To fulfill the promises that AI offers, we need unimaginable amounts of input data. Synthetaic will grow this high-quality data to unlock “impossible” AI.
