Cloud AutoML: From the research lab to the great outdoors

From the snow leopard to the giant panda, it’s no secret that some of the most beloved animals around the world are in danger of extinction.

Photos courtesy of the Zoological Society of London.

In fact, the average size of the world’s wildlife populations has declined by nearly 60% over the past 40 years. If these declines continue, many of the planet’s most iconic creatures could disappear completely within our lifetimes. But for organizations working to protect these animals, knowing where to start can be a challenge of its own: in order to defend at-risk species from human impact, conservationists must first identify which animals are in danger, what threat they’re facing, and where they’re headed next.

The Zoological Society of London, or ZSL—a conservation nonprofit dedicated to the protection of wildlife and their habitats—is taking a uniquely data-driven approach to this problem. They use camera traps to safeguard wildlife from poachers and monitor changes in wildlife population numbers. The cameras use motion and heat sensors to take a picture every time an animal or human passes. This process generates a large amount of data, which historically, ZSL has manually tagged and categorized, image-by-image. Categorizing imagery often takes months or even years to complete, and when you’re fighting to protect species under threat of extinction, time-consuming processes like this can be costly in more ways than one.

So ZSL teamed up with the Google Cloud team to test a new product, AutoML Vision (now in public alpha), which will allow them to translate this wealth of camera data into useful insights. Using AutoML, they’re developing custom machine learning models that can identify species within camera trap data and dramatically speed up large scale analysis. The testing is still ongoing, but the early results offer a signal of what’s possible when organizations are able to leverage the latest advancements in AI.

Turning moonshot research into product

While ZSL’s mission is somewhat unique, the challenges they faced before using AutoML Vision are remarkably common: the organization had plenty of information, but limited machine learning expertise to fully realize its potential. Last year, the Cloud AI team interviewed over 200 businesses and found that regardless of industry and organizational size, high costs and low accuracy were almost universal when it came to AI. For organizations fortunate enough to have machine learning resources in-house, highly skilled experts had to spend long, tedious hours manually training models for specific tasks. And for companies with more limited budgets, scenario-based, pre-trained models would have to suffice. Those pre-trained models were great for general classification (identifying a tiger versus an elephant), but companies like ZSL need specificity down to the sub-species level—something that could only be accomplished with high-fidelity, bespoke machine learning. "There has been a tension between ML tools that are easy to use, and ML tools that lead to high quality models. We have learned that in order to be successful, our customers need both,” explains Cloud AutoML Lead Product Manager Francisco Uribe.

In a parallel effort, the Google Brain team, a research group within Google AI, had begun research around the concept of “learning to learn.” Their goal was to create a neural network that could train another neural network without any outside assistance. A neural network is the software that actually does the “learning” in machine learning. To train a neural network, a developer provides the software with a detailed set of instructions describing the appropriate model to train, then trains it to recognize patterns within the dataset in a time-consuming, laborious process. A neural network that could train another neural network would take a significant burden off of researchers and engineers.

The concept of “learning to learn” isn’t a new concept in the field; researchers had been searching for a way to crack this elusive “meta-learning” problem as early as the late 80’s. In fact, one of our own researchers, Samy Bengio, wrote his PhD thesis in 1991 on “exploring the missing pieces of the puzzle of neural networks learning.” It wasn’t until the Brain team discovered and published their findings on reinforcement learning that “meta-learning,” or “AutoML,” as the team called it, went from theory to practice.

With these exciting breakthroughs, the Cloud AI team went to work translating this research into a product their customers could use. ZSL is using the first of these offerings, AutoML Vision, which focuses on image classification problems. Its drag-and-drop interface simplifies every step of the process—uploading images, tagging them, training the models and implementing the final product. This allows ZSL’s developers with limited machine learning expertise to translate their ideas into real, high quality, custom vision models. "We can train advanced models specific to different geographical regions and the species that live there. These models have the potential to help us categorize subspecies within a species group. For example, our experiments have shown models can go beyond ‘this is a monkey’ and start to identify subspecies within this group, such as a Red titi monkey versus an Emperor tamarin monkey. This level of fidelity is essential to our work," Sophie Maxwell, the Conservation Technology lead at ZSL explains.

Taking businesses to new heights

While AutoML Vision is still currently in public alpha, many of our enterprise customers are already seeing promising results. Organizations like ZSL are using AI to more quickly and accurately gather and analyze the data they collect. “In the future, it would be groundbreaking if conservation organizations could share and amass large scale data sets and use machine learning to give a health check of biodiversity across the planet and get a more real-time view of what's actually happening out there," Sophie Maxwell says. "We're nowhere near this yet, but we're taking initial steps to get us there."

AutoML Vision is just the first of many Cloud AutoML products to come, but it’s paved the way for future collaborations between our research and product development teams as we work on services for translation, object recognition, natural language processing, and more. And while we’re still early in our efforts, we’re excited to see what our customers continue to achieve with their new advanced AI capabilities. If ZSL is any indicator, they might just save the world.

Next Story

Powering the world with fusion