How AI Is Unearthing Hidden Scientific Knowledge | Sara Beery  ## Metadata - Author: [[TED Talks Daily]] - Full Title: How AI Is Unearthing Hidden Scientific Knowledge | Sara Beery - Category: #podcasts - URL: https://share.snipd.com/episode/fa78f1f5-e980-46c0-ae4f-4a2853ceb0fe ## Document Notes ## Highlights - **Most Life On Earth Is Unseen** - Scientists have observed only 2 million of an estimated 10 million species, leaving 80% of life unknown. - Sara Beery argues this massive unknown limits our ability to protect ecosystems and requires new approaches to knowledge discovery. Transcript: Sara Beery Imagine you're a doctor and you're trying to save the life of a patient, but you can only see a fifth of their body. How are you going to prescribe medicine? How are you going to do surgery? See, this is exactly the situation we're in with nature across the planet. We need to act now to protect ecosystems under threat, but there's so much we don't know about life on Earth. I'm an AI researcher and an ecologist. And as a professor at MIT, I lead a research group that develops methods to help us learn more about the natural world. And I see a future where AI can help exponentially increase our ecological knowledge across species and ecosystems. ([Time 0:03:19](https://share.snipd.com/snip/9c3ce8f7-aca5-40ee-9403-54c9ad71c5c2)) - **New Species Found Too Late** - The Tapenuli orangutan was discovered in 2017 and was already critically endangered at discovery. - Sara Beery uses this story to show species can go undescribed until it's nearly too late. Transcript: Sara Beery Take the Tapenuli orangutan. We discovered this orangutan in 2017. It's one of only three species of orangutan on Earth, and it was critically endangered before we even knew it existed. Traditional forms of data collection are just too slow to keep up with our current crisis. ([Time 0:06:27](https://share.snipd.com/snip/a39a41c4-63e0-49c2-b2a0-cda2ed3e8c8f)) - **Ecological Data Is An Untapped Goldmine** - Huge ecological databases like iNaturalist contain millions of labeled images with hidden extra information in the pixels. - Sara Beery calls these combined datasets an "ecological goldmine" that we haven't effectively mined yet. Transcript: Sara Beery Let's talk about just one of these databases, which is a platform called iNaturalist. 300 million images have been uploaded to this platform by passionate volunteers. In every single image, the community has identified a species, and that level of species occurrence data has already been transformative for science. But there is a hidden treasure trove of knowledge that remains in the pixels. In iNaturalist, this was labeled Grant's zebra. And it's clearly evidence that Grant's zebra were sighted in this place and time. But it shows us so much more than that. There are three Grant's zebra in this image. We can identify each of them to the individual level based on their unique stripe pattern. By identifying individuals, we can do things like monitoring how species move across the planet, looking at social networks of species, growth, health, even estimating the full Population size. These zebra are also coexisting with a herd of wildebeest. We can even see an oxpecker, a bird that eats ticks and helps reduce the spread of disease. We could look at the background of the image and identify the type and coverage of vegetation. We can estimate biomass, use that to learn about locally stored carbon. We can look at what the animals are eating in the image and build a stronger knowledge of a local food chain. Take this much knowledge in one image and multiply it by 300 million images in iNaturalist, and then add in our other ecological databases, millions of bioacoustic recordings in Xenocanto, Tens of millions of camera trap images in Wildlife Insights, thousands of hours of deep-sea footage in FathomNet. ([Time 0:06:58](https://share.snipd.com/snip/571508fd-4a98-4545-9c50-fc2ef6dd143b)) - **Automate Bulk Data Review** - Use AI to scan massive image collections because manual review would take decades of full-time work. - Sara Beery recommends automated methods to surface relevant data quickly instead of hand-labeling thousands of examples. Transcript: Sara Beery So say you want to look through all this data. Assuming it takes you about a second to look at every image, you would need to work full-time for 40 years to look through all the images in iNaturalist alone. And this is where AI is transformative. It can just help us look through all the data quickly. So an ecologist today, say they're interested in bird diets and they want to find examples of birds eating insects in the database, what they can do is they can train an AI model to help Them. So to do this, they collect hundreds or even thousands of examples to teach the model what to look for. Now, once they've trained this model, it's an incredible tool. It can very, very quickly find new examples of birds eating insects in the database. But this process of collecting hundreds or thousands of examples every ([Time 0:08:47](https://share.snipd.com/snip/924d57c3-1b2d-4a65-80b9-0452489eefe0)) - **Query Data With Language-Image Models** - Ask queries directly to databases with models that map scientific language to images instead of training new classifiers each time. - Sara Beery's team built Inquire to let ecologists search millions of images without collecting labeled examples or writing code. Transcript: Sara Beery This is what my team at MIT has been working towards, and we've developed a system that we call Inquire that helps ecologists find answers in the data without collecting any examples To teach an AI model or needing to write any lines of code. Now, under the hood, what we're doing is we're developing AI models that can learn and understand similarities between images and scientific language. And this is what allows us to just ask. So how does Inquire work? Well, first, an ecologist designs an experiment by taking a scientific question and breaking it down into a series of search terms that they can use to discover data that they'll analyze Downstream. So one of those terms might be bird eating insect. Now what happens is Inquire takes that search and it directly compares it to all 300 million images within seconds. It's engineered to do this both quickly and efficiently, which is important because it means the system is truly interactive, but it also requires far less computational power than A generative AI approach like ChatGPT. ([Time 0:10:10](https://share.snipd.com/snip/4f29eab0-0e58-4f72-acd9-24cc84765cdf)) - **Three-Hour Study Versus 1,560 Hours** - A collaborator used Inquire to find thousands of bird diet examples and analyzed seasonal diet shifts in about three hours. - Their results matched a manual study that required 1,560 hours of curation, demonstrating huge efficiency gains. Transcript: Sara Beery Now you have human verified examples of data that you can directly export and analyze. One of our collaborators used this system and they found thousands of examples of birds eating insects, but also seeds, fruit, nuts, carrion, nectar, plants. And then they took that data that they discovered quickly, and they analyzed differences in species diets between summer and winter. Now, what they found was that, yeah, some birds do eat insects in the winter. American robins actually do, but far less than they do in the summer. And some species, like American tree sparrow, that are incredibly dependent on insects as a a food source in the summer, don't eat them at all in the winter. This entire process, question to answer, took them about three hours. Another team spent 1,560 hours manually curating the data to do a similar study. And when you compare the results from Inquire to that study, you see an almost perfect match. ([Time 0:11:29](https://share.snipd.com/snip/799e0199-9204-4db3-a36d-7bbfe7802967)) - **Link Complementary Data Types** - Different ecological data types (images, sound, satellite, GPS) capture complementary perspectives on the same systems. - Sara Beery envisions linking these datasets to reveal hidden cross-data connections for richer ecological understanding. Transcript: Sara Beery And that brings up a whole new opportunity because all of these types of data are innately interrelated. They're all looking at the same thing. They're capturing complementary but distinct perspectives of life on Earth. And I can imagine a future where we have systems that help scientists quickly discover hidden connections between them all. Now, of course, this alone cannot solve our global nature crisis. But what it does do is it helps us maximize the value of data that we've already collected. And that means that then in turn, we can carefully understand what knowledge gaps remain and strategically use our resources to collect new data to fill those. ([Time 0:13:36](https://share.snipd.com/snip/2936ab6a-4f1e-43f6-abf9-3020319ae697)) - **Contribute To Public Databases** - Contribute observations to platforms like iNaturalist: every photo, sound, and note helps fill knowledge gaps. - Sara Beery urges public participation to build the complete picture of life on Earth for science and conservation. Transcript: Sara Beery And that is where all of you come in, because everyone can contribute. Everyone can collect data and upload it to platforms like iNaturalist. Every photo uploaded, every sound recorded, every observation shared is a piece of the puzzle. ([Time 0:15:23](https://share.snipd.com/snip/04de601d-b84f-47ae-a6d1-efba587dd14d))