Artificial Intelligence (AI) applications are growing at an unprecedented rate. Some observers even believe AI can solve some of the greatest problems facing humanity. However, there are not enough developers who know how to apply AI in sustainable ways.
To help combat this skills gap and to support sustainability goals, Arm teamed up with FruitPunch AI by sponsoring the AI for Bears Challenge. FruitPunch AI teaches applied AI with real-world challenges that address one of the UN’s 17 Sustainable Development Goals (SDG). Over 10 weeks, a global group of learners and experts tackled data problems faced by the challenge stakeholders. Once a team of 15-50 engineers is recruited, they participate in masterclasses, research, experimentation, implementation, and presentations.
In addition to my software engineering role at Arm, I volunteer for a non-profit I cofounded, the BearID Project. We develop AI applications for camera trap data to noninvasively monitor bears and other wildlife. One key reason bears are important is they are considered “umbrella” species. That is, they are wide ranging, and their protection also helps protect the many species that share their habitats. They face several conservation concerns, from habitat loss to climate change. The BearID Project addresses SDG #15: Life on Land across several conservation areas:
One of the main challenges faced by researchers like the BearID Project’s Conservation Scientist, Dr. Melanie Clapham, is processing vast amounts of images and videos. Networks of camera traps provide a means to monitor bear activity without disturbing them. However, it leads to thousands of hours of manual analysis. Due to the remote locations, data collection and data analysis, insights are often not available for many months or even years. While AI applications from the BearID Project are helping with analysis, near real-time solutions would enable more actionable data. For example, our co-stakeholder for the challenge, Hack the Planet, are working on a bear deterrent system to reduce human-bear conflict in Romania. For this, we need AI at the edge.
This is where Arm, FruitPunch AI and the AI for Bears Challenge comes in. Working with data sets from Hack the Planet and the BearID Project, the challenge had two goals:
A total of 19 AI for Good engineers spent more than 1500 hours over 10 weeks working across 4 groups:
Instead of focusing on the technical results of the AI for Bears challenge, I wanted to focus on the meaningful outcomes. If you are interested in the challenge results, you can read The Bear Necessity of AI in Conservation, which summarizes the final reports from each of the four teams. Here, I will focus on the learnings and impacts, both for wildlife as well as the people involved, affected by the challenge.
The challenge took 19 participants on a 10-week journey, learning about conservation, human-wildlife conflict, AI development flows and teamwork. On the latter point, Davide Coppola from the bear classification team said, “One key learning is that teamwork makes the dream work: everyone has different types of expertise and when put together for a common cause, great things can be achieved in record time.”
Several new tools were introduced to the teams. For example, the ML pipeline on edge team had access to Arm Virtual Hardware (AVH). Gaspard Bos, from the ML pipeline on edge team, pointed out that the virtual NXP i.MX93 development platform, “allowed us to get a head start in testing parts of our bear detection and identification pipeline on a virtual i.MX93 device and to get familiar with the platform.” Since the team was virtual and only one member had direct access to the physical device, AVH enabled the team to keep developing their application even when the board was not available.
At one point in the challenge, there was a request for the BearID Project to share more of their data. Dr. Clapham works closely with a First Nations organization, the Nanwakolas Council in British Columbia, Canada. Together they collect data on land under stewardship of the seven-member First Nations represented by the council. The BearID Project only shares this data under strict guidelines from the council. Gaspard had this takeaway on AI ethics: “I appreciate how the researchers respect the original custodians of the land and their involvement with the bear population. Their involvement and authority have implications for the practices of sharing data that might at first seem to complicate things for developers, but also force us to think twice about the data that we are handling and how to treat it with professional care.”
Arthur Caillau, who contributed to the bear detection & segmentation and bear identification teams, was already working with AI professionally and contributed to bear face detection and segmentation as well as bear identification. He said:
During this project, I refreshed my knowledge of metric learning to enhance bear face recognition...utilizing technology and engineering to address conservation challenges is incredibly rewarding.
Impact for wildlife conservation, especially bears, was a key goal. FruitPunch AI has contributed to this space before. "We are very excited this collaboration allowed us to use our community and experience in previous animal re-identification challenges to tackle this problem for bears as well and contribute to conservation efforts of these wonderful animals!", said FruitPunch AI CTO Sako Arts.
As noted in the challenge results blog posted by FruitPunch AI (see previous link), some of the AI models are working well. While not all the models were optimized to run on low-power edge devices, one of the bear-classifier models will be field tested this summer by Hack the Planet. It will be part of the second-generation bear deterrence system they are working on in Romania. Here is a short video about their current installation:
The challenge also provided a step toward on-device, near real-time solutions for monitoring. Utilizing the NXP i.MX93 development platform, the ML pipeline on edge team demonstrated classification, detection, and segmentation. Using the Arm Ethos-U65 Neural Processing Unit, the team showed 5-19 times speedup over CPU-based inference, which leads to significant power savings on a per inference basis. The next step is to integrate these models as a reference use case for smart wildlife cameras in the Software Defined Camera blueprint, an open source, cloud native, reference stack.
Besides the technical impact, there were also changes in attitude. For example, when asked how the challenge changed perception of wildlife for participants, Davide responded with “Regarding the human-bear conflict, it made me realize how the problem is bigger than I thought and how relatively simple solutions can have positive impacts to both sides of the conflict.” Other participants indicated they now have a better appreciation for the roles bears play in our world, and how our decisions affect them.
Some of the participants found more than technical learning, appreciation for wildlife and personal reward: they found a calling. Davide had this to say: “This challenge provided the opportunity to use my skills outside of my daily job and for a good cause. Ultimately, I hope this, and similar projects will put me on a trajectory of using AI for Good as main occupation.” Arthur was already leaning in this direction, saying “This challenge validated my decision to pivot my career toward conservation work.” During the challenge, Gaspard began working as an AI consultant for corporations and banks. He adds, “Even though this is the most commercial job I have ever had, I am still very impact driven so I mention this project in hopes of inspiring my colleagues and customers for doing something for the greater good.”
It wasn’t only the development teams that gained by participating in the challenge. Brian de Bart, team lead for system innovations at NXP had this to say, “The AI for Bears project was a great opportunity for NXP Semiconductors to contribute our microprocessor and AI technologies for a good cause. Our engineers got to work with an engaged team of AI enthusiasts and support them in building a solution for identifying bears in the wild. By deploying the developed AI models on the i.MX93 NPU, the team showed the effectiveness of the i.MX93 solution for achieving real-time classification and identification of bears on the edge”
Sponsoring the AI for Bears challenge was a fantastic way for Arm and our partners to help narrow the skills gap for AI developers while introducing them to Arm technology and addressing the sustainability goals. Participants learned the latest AI skills and developed a better appreciation for the natural world. If you are a developer interested in AI for Good, join one of the challenges hosted by FruitPunch AI. If you are a leader at a tech company, consider sponsoring a challenge or encouraging your teams to take part. Together, we can drive positive change for people and planet, unlocking sustainable growth and enabling progress on the United Nations’ Global Goals.
Learn more about Arm based AI technology applications
Learn more